Deep Learning Models for Accurate Leukemia & Lymphoma Detection

Deep Learning Models for Accurate Leukemia & Lymphoma Detection | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep Learning Models for Accurate Leukemia & Lymphoma Detection Benedict Onochie Ibe, Dagogo Godwin Orifama, Ali Dan, Ikechukwu Nwagbo Enumah, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8155029/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Acute lymphoblastic leukemia (ALL) and Lymphoma are important diseases that need to be detected early, but detection with traditional methods may be slow and inconsistent. Although current AI methods present potential advantages, they are frequently limited by their reliance on small input data, overfitting, and a lack of external verification, which are unfavorable for their clinical implementation. This paper proposes a unified framework for creating a shared deep learning model for dual disease detection, based on the ConvNeXt design. ConvNeXt's superiority was demonstrated in a first ablation, wherein the Model was compared to Swin Transformer, ResNet101, VGG16, and a custom CNN on an ALL dataset, achieving a best accuracy of 99.69%. The selected ConvNeXt model was then optimized and retrained on a larger dataset consisting of both ALL and Lymphoma samples. As a result, there was only one highly consistent diagnostic tool that attained an accuracy of 99.72% on the mixed test set. More importantly, the practical utility of the framework in obtaining significant results was validated through extensive testing using completely new, previously unseen external datasets. The framework demonstrated an outstanding degree of generalizability, achieving 100% accuracy on an independent ALL dataset and 97.73% on an independent Lymphoma dataset. This work presents a well-represented, fully automated, and externally validated hematological diagnosis system that demonstrates a possible route of implementing trusted AI in direct care practice. Artificial Intelligence and Machine Learning Acute Lymphoblastic Leukemia (ALL) Deep Learning Peripheral Blood Smear (PBS) Lymphoma Convolutional Neural Networks (CNNs) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction Hematologic malignancies like Acute Lymphoblastic Leukemia (ALL) and lymphoma have been diagnosed and detected with better methods over the last century. Initially, the diagnosis was made based on clinical findings and routine histological examination. The progression of imaging modalities from simple X-ray and lymphography to sophisticated CT, FDG-PET, and PET/CT has significantly enhanced the sensitivity and specificity of lymphoma staging and assessment of treatment outcomes [1]. Additionally, in the case of ALL, the combination of molecular diagnostics, genetic tests, and MRD detection using flow cytometry and PCR has altered the approach to risk categorization and disease management [2] [3]. These advancements have not only increased survival rates but also facilitated the development of personalized treatment plans, particularly with the aid of immunotherapies targeted at specific genetic mutations. Nevertheless, there are still issues regarding the improvement of these diagnostic methods for clinical practice. In the diagnosis of lymphoma, PET/CT fusion is superior to other modalities; however, variability in methods and high costs have limited its widespread use. Likewise, the use of MRD markers in ALL has been found to help detect early relapse; however, the role of these biomarkers in various patient populations, including adults, is currently under further investigation [4]. New technologies, such as radiomics and deep learning in imaging, as well as next-generation sequencing for genetic profiling, are showing potential in enhancing diagnostic accuracy and personalized care in cancer management [5]. Nevertheless, there is a need to establish the validity of these new technologies and to eliminate the existing differences in diagnostic care between different healthcare settings. Emerging architectures such as Vision Transformers (ViTs) and hybrid models like ConvNext offer promising solutions. ViTs utilize self-attention mechanisms to capture intricate relationships across different image regions, outperforming CNNs in haematological image classification tasks. For instance, ViTs achieved an accuracy of 98% [6], compared to 92.27% at a learning rate of 0.003 [7] for CNNs, highlighting their potential in enhancing diagnostic accuracy for ALL. Similarly, ConvNeXT models combine convolutional layers with attention mechanisms, optimizing feature extraction and performance. Despite these advancements, challenges remain in deploying these models in clinical settings, particularly in resource-constrained environments. Lightweight architectures, such as MobileNet and SqueezeNet, provide a balance between accuracy and computational efficiency, making them suitable for rapid diagnostics. However, their performance is generally lower than that of more complex models. Pre-trained transfer models, including ResNet101, VGG16, and EfficientNet [8], have demonstrated high accuracy due to their rich feature representations. However, the need for extensive computational resources and large labelled datasets limits their widespread clinical application. Therefore, there is a critical need for a diagnostic model that integrates the strengths of advanced deep learning architectures, such as the accuracy of ViTs, the hybrid efficiency of ConvNeXT, and the computational economy of lightweight models, while minimizing their limitations. This study addresses this gap by developing an optimized diagnostic framework specifically tailored for the detection and classification of ALL, ultimately aiming to enhance diagnostic accuracy, efficiency, and accessibility in diverse clinical settings. 1.1. Objectives of the Study To develop a unified deep learning framework for the dual detection of ALL and Lymphoma To validate the proposed ConvNeXt-based framework through an ablation study against other prominent architectures. To assess the framework's generalizability on independent, unseen external datasets. Build one single Model to detect or predict the different stages of Acute Lymphoblastic Leukemia (Benign, Early, Pre, and Pro), and the different classes of Lymphoma (chronic lymphocytic leukemia, follicular lymphoma, and mantle cell lymphoma), thereby creating one Model to detect two types of ALL and Lymphoma. 2. Literature Review 2.1. Advancements in Acute Lymphoblastic Leukemia (ALL) Detection Acute Lymphoblastic Leukemia (ALL) is a type of cancer that affects the lymphoid cells. Conventional methods of diagnosis are based on blood smear analysis and immunophenotyping, which are time-consuming and dependent on the availability of pathologists. However, with the development of AI and deep learning, several innovative strategies have been proposed to enhance the diagnostic accuracy and speed of ALL. Deep learning models, especially Convolutional Neural Networks (CNNs), have been applied to the detection of ALL from microscopic images. A study using ResNet-50, VGG-16, and custom convolutional networks achieved the highest validation accuracy of 84.62%, highlighting the potential of AI in automated leukemia diagnosis [9]. Other models have attempted hybrid approaches, such as combining CNN with SVM, which achieves nearly 100% accuracy in some cases [10]. In addition, fluorescence spectral methods for non-invasive ALL diagnosis have 88% sensitivity and 80% specificity, offering a new approach to distinguish leukemia from standard samples [11]. Automated detection techniques using YOLOv4 have also been introduced, achieving a mean average precision (mAP) of 98.7%, proving the feasibility of object detection models for blast cell identification [12]. Transfer learning techniques have been applied to histopathological images, resulting in a significant improvement in classification accuracy [13]. However, challenges remain in standardizing datasets and ensuring the generalizability of models across diverse populations. [14] Investigates the application of deep learning models for leukemia classification with the aid of AI, specifically DenseNet201, EfficientNet B2, and ConvNeXT Small. Of the three models, the ConvNeXT model had the highest accuracy of 92.88%, sensitivity of 83.89%, specificity of 97.07%, and F1-score of 88.22%. These results indicate that the ConvNeXT model is more effective at identifying leukemia cases and distinguishing between subtypes compared to the other two models. Hence, the proposed model holds potential for assisting clinicians in making early and accurate diagnoses of leukemia, which is crucial for managing this deadly disease. Research [15] has been conducted on various deep learning models for leukemia classification, including ConvNeXt, DenseNet201, EfficientNet B2, and ensemble models. Accuracy was highest for ConvNeXt at 92.8%, making it the best architecture for detecting leukemia. The study also highlights the significance of early identification and the feasibility of AI-based solutions for medical diagnosis. 2.2. Advancements in Lymphoma Detection Lymphoma, a cancer of the lymphatic system, has also been diagnosed more accurately with the help of AI-driven approaches. The traditional methods include whole-body PET/CT imaging and histopathological examination, but these are often time-consuming and prone to interobserver variability. AI has been integrated with lymphoma detection and deep learning-based automated PET/CT analysis, which achieved an 85% accurate positive rate (TPR) using CNNs [16]. Histopathological image analysis using CNNs has also been effective in lymphoma subtype classification, with EfficientNet-based models reaching 95.56% accuracy in distinguishing different NHL subtypes [17]. Faster R-CNN models have been trained on blood cell datasets, yielding over 96% detection rates for lymphoma cells [18]. Moreover, a systematic review on deep learning in PET image interpretation found AI models achieving an AUC of 0.963 in lymphoma detection [19]. However, there are still issues with the size of the datasets, the absence of external validation, and the necessity to translate the models for real clinical use. Although AI models are accurate in diagnosis, their integration into regular clinical practice and comparison with human pathologists has not been fully understood [20]. This study [21] aims to compare the preoperative differentiation between Primary Central Nervous System Lymphoma (PCNSL) and Glioblastoma (GBM) using deep learning and radiomics-based AI models. Three models were evaluated: The Radiomics model (AUC = 0.86), the ConvNeXt model (AUC = 0.89), and the MobileVIT Model (AUC = 0.91). The Max-Fusion Model, which combines radiomics and deep learning, achieved the best performance with an AUC of 0.92. This fusion-based approach significantly enhances tumor classification accuracy and may thus aid clinicians in making non-invasive decisions. The model that performed the best was the Max-Fusion model, which combines radiomics and deep learning, with an AUC of 0.92. This fusion-based approach significantly improves tumor classification accuracy and, in turn, may be helpful to clinicians in making non-invasive decisions. 2.3. Problems However, several problems persist. Many of the models are trained on the institution-specific small datasets, which are not generalizable. Furthermore, there is no standardized evaluation framework for such tasks, making it difficult to compare different approaches. The lack of external validation is also a concern that raises questions about the real-world applicability of these models. The aforementioned challenges include overfitting and the need for high computational resources, both of which are discouraging for widespread clinical adoption. The problems or challenges are explained in more detail in Table 1 below: Table 1 Problems identified in previous research about the implementation of Artificial Intelligence and Deep learning Problem/Gaps Identified Validation from Research Sources Lack of Large-Scale Datasets - Many AI models are trained on small, institution-specific datasets, which limits generalizability. Source: A systematic review found that many studies lacked large datasets and external validation, which limited their clinical application [22]. Also, the PET/CT lymphoma detection study [16] highlighted that training CNNs on limited datasets affects performance Limited External Validation - Many models are trained and tested on the same dataset, without independent validation. Source: The review of AI in PET imaging highlighted the lack of external validation in lymphoma detection models [19]. Additionally, the histopathological AI model study mentioned that it had limited testing on diverse datasets [17]. Overfitting in Deep Learning Models - Some AI models achieve near-perfect accuracy on training data but may not generalize to real-world cases. Source: The YOLOv4 leukemia detection study mentioned that deep learning models require careful tuning to avoid overfitting [12]. Also, the ALL deep learning comparison highlighted concerns about model robustness [9]. Computational Resource Requirements - Many deep learning models require expensive GPUs and substantial computing power. Source: The histopathological AI study mentioned computational requirements as a challenge for real-world adoption [17]. Additionally, studies utilizing deep CNNs have emphasized the need for specialized hardware [9]. 2.4. Proposed Solution For The Research Gather data from several sources to build a big and diverse dataset to increase the model's generalization for ALL and lymphoma detection. Enhance model generalization with external validation, which ensures that the models are trained on one dataset and validated on completely different datasets that have not been seen before for real-world applications. Data augmentation, cross-validation, and regularization are employed to enhance the model's robustness and prevent overfitting. Ablation or comparative studies using Transformers (SWIN), ConvNeXT, and CNN-based architectures for high accuracy, efficiency, and computational optimization. Design a single AI system that can identify various stages of ALL (Benign, Early, Pre, Pro) and types of lymphoma (CLL, FL, MCL). To achieve these objectives, we propose a framework centered on the ConvNeXt architecture, whose hybrid design is uniquely suited to overcome the feature extraction challenges in hematological imaging. 3. Research Methodology The initial phase of this study focused on detecting Acute Lymphoblastic Leukaemia (ALL) and comparing the effectiveness of different deep learning models on an ALL dataset. The images were taken from Kaggle [23] and included four stages of ALL: Benign, Early, Pre, and Pro. The images were available in two forms: The original and the segmented, to see how the feature extraction differs. The images were imported into a pandas DataFrame, named by class labels and image type, and EDAs were conducted on the data. To identify imbalances, a class distribution analysis was conducted, and sample images from each class were displayed to gain an understanding of the dataset's structure. 3.1. Data Preprocessing Several preprocessing steps were undertaken to standardize the input data: Path Setup and Data Loading: The dataset's structure was also explored, and a DataFrame was generated to store the image paths, class labels, and image types. Label Encoding: Both ALL and Lymphoma class labels were numerically valued in the following: ALL Classes: Benign (0), Early (1), Pre (2), Pro (3) Lymphoma Classes: CLL (4), FL (5), MCL (6) Data Splitting: The dataset was divided into training (80%), validation (10%), and test (10%) sets, created using stratified sampling to maintain class balance. Image Preprocessing Function: A specific function was written to import and resize the images to 224x224 pixels for uniformity with other networks. This function also undecoded and denormalized the images to match the chosen architectures. Shuffling and Batching: The training, validation, and test sets were randomly shuffled to introduce some element of randomness and to mitigate any potential bias during training. An interval of 32 was chosen for the batch size to optimize memory utilization and model training time. Class Balancing: To check for any class imbalance, the dataset was grouped by class labels, and the distribution of images per class was visualized using bar charts. This step was crucial to ensure that each class had a representative number of samples, which informed the class balancing techniques employed later in the preprocessing pipeline, as shown in Fig. 1 (a) below. The dataset class imbalance was therefore addressed using oversampling, which was applied through resampling techniques to ensure uniform representation across classes, as shown in Fig. 1 (b). 3.2. Data Augmentation To enhance the robustness of the models, data augmentation strategies were used during the training process: Random Flipping : To create different views, images were flipped horizontally. Random Rotation : A small amount of random degree rotation (up to 20%) was applied to the images to account for variation in orientation. Normalization : All pixel values were normalized to the range of 0 to 1 to guarantee stable training. After the preprocessing steps, we conducted an initial comparative analysis of multiple deep learning models for ALL detection, comparing the performance of the Swin model, ResNet101, VGG16, a Custom CNN model, and the ConvNeXt model. It was observed that ConvNeXt achieved the highest accuracy. However, to extend the applicability of the model beyond ALL detection, the dataset was expanded to include Lymphoma. This inclusion aimed to create a comprehensive haematological model using the ConvNeXt model due to the fact that it performed best in the initial comparative analysis, capable of detecting both ALL and Lymphoma, thereby improving diagnostic utility. The Lymphoma dataset was downloaded from Kaggle [24] using the Kaggle API, specifically from the Malignant Lymphoma Classification dataset. The dataset consisted of images labelled under three primary lymphoma classes: Chronic Lymphocytic Leukaemia (CLL), Follicular Lymphoma (FL), and Mantle Cell Lymphoma (MCL). The Lymphoma images were preprocessed similarly to the ALL dataset, and TIFF images were converted to JPG format for consistency. To ensure model generalization and prevent bias, the ALL and Lymphoma datasets were merged into a single dataset, resulting in a combined dataset containing images from seven classes (four ALL classes and three Lymphoma classes). The dataset was then checked for class imbalance, and to address any disparities, oversampling techniques were applied using resampling to balance the classes before training. To ensure consistency in data formats, the Lymphoma dataset’s TIFF images were converted to JPG. The ALL and Lymphoma datasets were then merged into a single dataset, resulting in a seven-class classification problem (four ALL classes + three Lymphoma classes). To further validate the robustness of the model, an additional Lymphoma dataset from Kaggle [25] and an additional ALL dataset from Roboflow in Pascal VOC format were downloaded to serve as additional external test datasets. These datasets were used to test the model on unseen data. 3.3. Model Selection and Proposed Architecture Following dataset preparation and preprocessing, we examined a few of the best contemporary models to establish the best architecture for our intended diagnostic system. Based on this analysis, the ConvNeXt model would be our proposed method due to its novel hybrid structure. This selection was benchmarked and validated by other models (Swin Transformer, ResNet101, VGG16, and a baseline CNN) as part of an ablation study. 3.3.1. ConvNeXt Model ConvNeXt is a recent convolutional model designed with a transformer-like architecture to enhance the efficiency of convolutional operations. It introduces depthwise convolutions, LayerNorm, and GELU activation, which enhance feature extraction. ConvNext is a modern convolutional architecture trained with pre-trained weights, where ConvNext was fine-tuned for the task of imaging leukemia stages and assigning them the correct class, based on both local and global features. The conventional convolution operation is given by (1) below: $$\:y\:=\:x\:+\:f\left(x\right)$$ 1 Where f(x) is made up of: Depthwise Conv (kernel size 7×7) Layer Normalization Pointwise Conv (1×1) expanding channels by 4 GELU activation Pointwise Conv (1×1) reducing channels. Layer Scale (learned per-channel scaling) And then fully expressed mathematically as (2): $$\:f\left(x\right)=\:{\text{P}\text{W}\text{C}\text{o}\text{n}\text{v}}_{\text{d}\text{o}\text{w}\text{n}}\left(GELU\left({PWConv}_{up}\left(LN\left(DWConv\left(x\right)\right)\right)\right)\right)⊚\:\lambda\:$$ 2 Where \(\:\lambda\:\) is a learnable parameter for layer scaling. 3.3.2. SWIN Transformer Model The transformer-based architecture used in this paper was initialized with pretrained weights and fine-tuned on the ALL dataset. The model capacity was enhanced through attention mechanisms to leverage the ability to model long-range dependencies for the classification of leukemia stages. The Swin Transformer is designed with a hierarchical architecture, where an image is treated as a sequence of non-overlapping windows to model both local and global dependencies powerfully. Such shifting window mechanisms are effective in capturing fine details from small regions while providing a broader perception of the whole image. While its multi-head self-attention inside the window enabled robust feature extraction, it likely endowed the model with high accuracy in the classification of all stages. Its mathematical representation involves: Patch Partitioning : The input image is divided into non-overlapping patches as seen in (3). $$\:\text{P}\text{a}\text{t}\text{c}\text{h}\:\text{s}\text{i}\text{z}\text{e}\left(\:=\:P\right),\hspace{1em}\text{w}\text{h}\text{e}\text{r}\text{e}\:P\:\text{i}\text{s}\:\text{t}\text{h}\text{e}\:\text{p}\text{a}\text{t}\text{c}\text{h}\:\text{d}\text{i}\text{m}\text{e}\text{n}\text{s}\text{i}\text{o}\text{n}.$$ 3 Linear Embedding : Each patch is linearly projected into a feature vector, where \(\:{W}_{p}\) is the learnable weight, \(\:{x}_{p}\) is the Input patch, \(\:{and\:b}_{p}\) is the learnable bias as seen in (4). $$\:{z}_{p}={W}_{p}\cdot\:{x}_{p}+{b}_{p}$$ 4 Shifted Window Attention : Self-attention is computed within local windows, and windows are shifted between layers to enable cross-window connections, and is represented in (5) below: $$\:\text{Attention}\left(\varvec{Q},\:\varvec{K},\:\varvec{V}\right)=\text{Softmax}\left(\frac{\varvec{Q}{\varvec{K}}^{\varvec{T}}}{\sqrt{{d}_{k}}}\right)\varvec{V}$$ 5 Where Q, K, V are the Query, Key, and Value matrices, and \(\:{d}_{k}\) is the Scaling factor (dimension of keys) 3.3.3. ResNet101 Model ResNet101 is a deep convolutional neural network architecture designed by Microsoft Research in 2015 [26]. It is noted for having a very high number of layers (101 in total) and yet preventing the vanishing gradient problem through the use of innovative residual connections. The main idea of ResNet101 is the residual block, where instead of learning unreferenced functions, each layer learns residual functions that depend on the layer inputs. These "skip connections" enable the training of significantly deeper networks than previously possible. The idea was built on the success of ResNet, which won several visual recognition challenges in 2015. The architecture comprises a single convolutional layer followed by 33 bottleneck blocks (each containing three convolutional layers), organized into four stages of increasing feature map depth (64, 128, 256, and 512 channels). The bottleneck design is implemented to reduce and then restore dimensions using 1×1 convolutions, with a 3×3 convolution in between to preserve dimensions while efficiently managing computational resources. 3.3.4. VGG16 VGG16 is a convolutional neural network architecture developed by the Visual Geometry Group at Oxford University in 2014 [27] and characterized by its simplicity and depth, specifically comprising 16 weight layers (13 convolutional layers and three fully connected layers). The architecture is characterized by a consistent pattern of 3x3 convolutional layers with ReLU activation functions, followed by max-pooling layers for down-sampling. The number of filters is increased progressively (from 64 in early layers to 512 in deeper layers) while the spatial dimensions are reduced through pooling. One of the design philosophies that proved to be highly effective was the use of multiple small filters rather than fewer large ones. Transfer learning was applied using ResNet-101 and VGG-16, which were pre-trained on the ImageNet dataset. These models were modified by replacing the fully connected layers with custom classifiers tailored to the ALL dataset. The initial layers were frozen to retain general image features, and the later layers were fine-tuned for our task. 3.3.5. Custom CNN Model A custom CNN architecture was built with several convolutional layers followed by max-pooling and fully connected layers. This model served as a baseline to compare the effectiveness of transfer learning and transformer-based models. After evaluating these models on all detection tasks, the ConvNeXt model emerged as the best-performing architecture, achieving the highest accuracy across all tasks. Given this, the study was extended to detect both ALL and Lymphoma by training the ConvNeXt model on the expanded dataset, which is a combination of the ALL dataset and the Lymphoma dataset. 3.3.6. Evaluation Metrics Accuracy : Accuracy is defined as the percentage of the models' predictions (including both true positives and true negatives) that match the total number of cases. It is stated as: "What percentage of all predictions did the model get right?" However, accuracy can be misleading on imbalanced datasets where one class is significantly outnumbered by the other. Accuracy is represented mathematically as (6) below: $$\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\:\frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{T}\text{N}+\text{F}\text{P}+\text{F}\text{N}}$$ 6 Precision : Precision is the percentage of correct identifications. It is stated as: "When the model outputs the positive class, what is the chance of being accurate?" Precision is stated as the fraction of the true positives over the sum of the true and false positives. It is essential in cases where false positives are expensive, and is expressed as (7) below: $$\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\:\frac{TP}{TP\:+\:FP}$$ 7 Recall : Recall (also known as sensitivity) measures the proportion of actual positives that are correctly identified. It answers: "What percentage of actual positive cases did the model identify?" Recall is calculated as true positives divided by the sum of true positives and false negatives. It is critical in scenarios where missing a positive case has serious consequences, such as disease detection. Recall is expressed mathematically by (8): $$\:Recall\:=\:\frac{TP}{TP\:+\:FN}$$ 8 F1 Score The F1 score is the harmonic mean between precision and recall, which is a single value that considers both of them. It is on a scale of 0 to 1 and is especially useful when the goal is to find a tradeoff between precision and recall or when the classes are imbalanced. The F1 score is calculated as (9) $$\:\text{F}1-\text{S}\text{c}\text{o}\text{r}\text{e}\:=\:2\:\times\:\:\frac{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\times\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}$$ 9 Where TP is the True Positives, TN is the True Negatives, FP is the False Positives, And FN is the False Negatives 3.3.7. Training ConvNeXt for ALL and Lymphoma Detection To train the final model on the combined dataset, the following configurations were used: Model Architecture : ConvNeXt, initialized with pre-trained weights, was fine-tuned for multi-class classification across the seven ALL and Lymphoma classes. Loss Function : Categorical cross-entropy loss was applied. Optimizer : Adam optimizer (lr = 0.0001) with a scheduler to reduce learning rate on plateau. Training Configuration : The model was trained for 10 epochs with early stopping based on validation loss. Evaluation Metrics : Accuracy, precision, recall, and F1-score were used for assessment, alongside confusion matrices to visualize classification performance. 3.3.8. External Dataset Evaluation After training the comprehensive ConvNeXt model, its efficacy was tested on external datasets that were not part of the training process. To validate the model's robustness, two additional datasets were downloaded: 1. A different Lymphoma dataset from Kaggle - This dataset was downloaded using the Kaggle API and was used to test how well the model generalizes to unseen Lymphoma cases. 2. A different ALL dataset from Roboflow - This dataset served as a secondary validation set for assessing the model's performance on new ALL samples. The trained ConvNeXt model was evaluated on both external datasets, and accuracy metrics were generated to determine its performance. This step ensured that the model was not overfitting to the original dataset and was capable of detecting both ALL and Lymphoma in diverse datasets. By expanding the dataset and retraining the model, this research significantly enhances hematological diagnostics, enabling broader leukaemia and lymphoma detection using deep learning. The methodology outlined ensures full reproducibility for further research in leukemia and lymphoma classification. A schematic of the complete research methodology is presented in Fig. 2 . 4. RESULTS AND DISCUSSION The test results of the models revealed that the ConvNeXt model performed the best, achieving an impressive 99.69% accuracy, followed closely by the Swin model with an accuracy of 99.39%. Both models demonstrated superior performance in differentiating between the stages of ALL, likely due to their modern architecture design, which efficiently captured both local and global features within the images. ConvNeXt Model (99.69%) This model performed exceptionally well due to its ability to capture spatial hierarchies and visual features efficiently. ConvNeXt's modern convolutional structure is designed to mimic the strengths of both convolutional and transformer models, allowing it to achieve state-of-the-art performance. The training accuracy and loss curves are shown in Fig. 3 (a). At the same time, the ConvNeXt model demonstrated high precision, with only a single misclassification of a 'Benign' sample as 'Pre' (Fig. 3 b) Swin Model (99.39%) Training the SWIN transformer model obtained a very high accuracy of 99.39% which shows its efficacy in perfectly classifying ALL classes, which is proof that the balance of Swin is done between the tradeoff of accuracy and computational efficiency, more so for this medical image task. The training accuracy and loss curves are shown in Fig. 4 (a). In contrast, the confusion matrix for the Swin vision transformer model, which showed it only misclassified the 'Benign' label as 'Early', is shown in Fig. 4 (b) below ResNet101 (98.47%) As expected, ResNet101, a deep residual network, also performed well, given its depth and ability to learn hierarchical features. However, its accuracy was slightly lower than ConvNeXt and Swin model, possibly due to overfitting on certain leukemia stages. The confusion matrix for the ResNet101 model, seen in Fig. 5 (b), showed it misclassified the 'Early' label as 'Benign' the most (2) due to the very minimal difference between the two stages, while the training accuracy and loss can be seen in Fig. 5 (a) VGG16 (96.32%) While VGG16 is known for its strong feature extraction capabilities, its performance lagged behind more modern architectures like ConvNeXt and ResNet101. This may be due to the simplicity of the VGG architecture, which does not incorporate the advanced residual or attention mechanisms found in newer models. The confusion matrix for the VGG16 model, shown in Fig. 6 (b), indicated that it misclassified the 'Benign' label as 'Early' the most (9). At the same time, the training accuracy and loss for the VGG16 are seen in Fig. 6 (a). Custom ALL Model (90.80%) The combination model, which integrates features from multiple architectures, performed reasonably well but did not match the precision of individual models, such as ConvNeXt or Swin model. This lower accuracy may reflect the challenge of combining features from multiple models without over-complicating the learning process. The training accuracy and loss of the Custom ALL model is seen in Fig. 7 (a), while the confusion matrix for the ResNet101 model, which showed it misclassified the 'Early' label as 'Pre' and 'Early' as 'Benign' the most (6) and is shown in Fig. 7 (b) below. 4.1. Confusion Matrix Insights The confusion matrices generated for each model provided additional insights into the classification performance. The ConvNeXt model and Swin model both showed minimal misclassification across the stages of ALL. The errors that did occur were mainly between adjacent stages, which is understandable given the visual similarity between these stages. For ResNet101 and VGG16, misclassifications were more frequent, particularly between the "Early", "Benign", and "Pre" stages. This suggests that while these models are powerful, they may struggle with finer-grained distinctions between certain stages of ALL, particularly when features are subtle. The implementation of data augmentation (random flips and rotations) and normalization was instrumental in improving model generalization and preventing overfitting. Augmentation allowed the models to train on a more diverse dataset, leading to better accuracy across models. The normalization of pixel values ensured smoother convergence during training, especially for deeper models like ResNet101. The complete comparison of the trained models on only the ALL dataset is summarized in Table 2 below: Table 2 Table 2 shows our first ablation study on the Acute Lymphoblastic Leukemia (ALL) dataset. The table gives a comparative overview of the accuracy of the performance and precision, recall, and F1 score of each of the tested architectures, supporting the choice of the ConvNeXt model Model Accuracy (%) Precision Recall F1 Score ConvNeXt model 99.69 1.0 0.99 1.0 Swin model 99.39 1.0 0.99 1.0 ResNet101 98.47 1.0 0.99 1.0 VGG16 96.32 0.96 0.96 0.95 All model 90.80 0.95 0.96 0.95 4.2. Discussion on Model Architectures The ConvNeXt model excelled in this study due to its modern architecture, which effectively combines convolutional principles with the strengths of vision transformers, allowing it to recognize subtle variations in stages of Acute Lymphoblastic Leukemia (ALL). The Swin model, with its design, achieved near state-of-the-art performance, demonstrating that efficiency can coexist with complexity, making it suitable for resource-constrained clinical settings. Meanwhile, both ResNet101 and VGG16, despite being effective transfer learning models, indicated that further fine-tuning is necessary when adapting models pretrained on general datasets like ImageNet for specific medical imaging tasks; ResNet101 offered refined feature extraction but still lagged behind ConvNeXt and Swin. Lastly, the ALL model, which aimed to integrate features from various architectures, underperformed due to the under-complication of the learning process and ineffective feature integration. Performance Of The ConvNeXt Model on Extended Dataset (ALL + Lymphoma) Based on the previous evaluations, ConvNeXt was chosen as the final model to be trained on the extended dataset that includes both ALL and Lymphoma cases. The reason for choosing ConvNeXt, a hybrid CNN-Transformer architecture, is that it is capable of learning hierarchical spatial features and also leverages the attention mechanism for enhanced feature extraction. The training configuration used while training the ConvNeXt model on the extended dataset is given below • Input Size : 224 × 224 • Batch Size: 32 • Optimizer : Adam (lr = 0.001) • Loss Function : Cross-Entropy Loss • Training Epochs: 10 During training, the ConvNeXt model converged rapidly due to its pre-trained weights and new architectural enhancements, which improved the efficiency of convolutional operations. To enhance generalization, random flipping and rotation were also used as data augmentation techniques. The training and validation accuracy and loss curve can be seen in Fig. 8 (a) below: The validation accuracy was consistently above the training accuracy, which means that the model did not overfit. The new ConvNext model, which was finally trained, was quite impressive, with an accuracy of 99.72% on the test set, proving its efficacy in detecting both ALL and Lymphoma and classifying them correctly. To gauge the model's classification performance, a confusion matrix was created in Fig. 8 (b) to show how the model performed in correctly classifying all the disease categories where we see that only 1 Pro ALL class was misclassified as CLL. External Dataset Evaluation of the Best Performing ConvNeXt Model The results from testing the ConvNeXt (ALL + Lymphoma) model on a completely unseen external dataset reflect the generalization ability and robustness of the model. The model reached an excellent accuracy of 97.73% on the external Lymphoma dataset. The model also got 100% accuracy on the external ALL dataset which shows that it has learned the essential features of ALL subtypes and is capable of classifying new cases. The high accuracy in both datasets shows that the model does not overfit the training data but rather generalizes the features. The outstanding performance of the model reveals its clinical significance, and thus it can be used as a reliable AI-based diagnostic tool for distinguishing between ALL and Lymphoma in real clinical practice. The confusion matrix, which shows the detailed performance of the ConvNeXt (ALL + Lymphoma) model on the ALL-external dataset, can be seen in Fig. 9 (b), while the confusion matrix for the test of ConvNeXt (ALL + Lymphoma) on the external Lymphoma data is seen in Fig. 9 (a) Table 3 Table 3 below summarizes the accuracy performance of the ConvNext (ALL + Lymphoma) model on both the extended and hybrid datasets and the two external datasets. Dataset Accuracy (%) ALL + Lymphoma 99.72 External ALL Dataset 100.00 External Lymphoma Dataset 97.73 Table 4 Comparison of the performance of different state-of-the-art (SOTA) models from varying research work with results from this research work, which details how our model and approach outperformed previous studies/research. Authors Dataset Model Accuracy (%) Description This study ALL + Lymphoma data ConvNeXt 99.72 - [28] normal vs. malignant cells (C-NMC) dataset CNN 91.10 Utilized CNN for ALL classification [29] WBC datasets Resnet-50 90.00 Classification of Activated and Inactivated Neutrophils for Rapid Immune State Monitoring [30] ALL data VGG16 98.15 Utilized VGG16 for ALL classification [31] Blood samples Swin Transformer 96.80 Hematology analysis framework using optical diffraction tomography [32] Lymphoma Content-based image retrieval system 98.74 Incorporation of deep learning with a traditional learning approach [33] ALL ALNet 94.50 A sequential CNN-based system to predict the initial diagnosis of acute leukaemia [34] ALL XGBoost 99.31 Efficacy of various machine learning models in the prediction of Acute Lymphoblastic Leukemia (ALL) [35] Lymphoma BNN 99.00 BNN for follicular lymphoma detection [36] Whole slide images - Lymphoma DLCNN 88.00 To investigate a deep learning convolutional neural network (DLCNN) for computer-aided detection of MCL [37] Lymphoma CNN 95.00 Use Deep Learning with a convolutional neural network (CNN) algorithm to build a lymphoma diagnostic model All Model 90.80 0.95 0.96 0.95 To illustrate the performance of our suggested framework, we compared our results with some state-of-the-art (SOTA) approaches from the recent literature (Table 4 ). The performance of our unified ConvNeXt model, based on the compound dataset (99.72% accuracy), is also better or at least competitive with that of specialized models specific to ALL or lymphoma. For example, our model outperforms the CNN model by [28] (91.10%) and the ALNet system [33] (94.50%) in detecting ALL. Our single-model-based two-disease detection scheme has been proven effective and innovative on independent datasets. Conclusion and Recommendations This research has effectively developed a sophisticated AI baseline diagnostic framework for detecting and predicting Acute Lymphoblastic Leukemia (ALL) and Lymphoma using a single model within deep learning frameworks. A comparative analysis of multiple models was initially performed, and the ConvNeXt model was selected as the most efficient among Swin Transformers, ResNet101, VGG16, and a Custom CNN, achieving an accuracy of 99.69% on the ALL dataset. The ConvNext model was then used to train a multiple dataset comprising both the ALL and Lymphoma datasets, where it achieved 99.72% on the extended ALL + Lymphoma dataset. The model was tested on external, unseen datasets, where it achieved 100% accuracy on the ALL dataset and 97.73% accuracy on the Lymphoma dataset, demonstrating the model's generality and potential for real-life use. These results support the idea that combination architectures like ConvNeXt that incorporate both convolutional and transformer-based features offer better feature learning and classification performance in the field of hematological image comprehension. The study also addressed important issues in the application of AI in medical diagnostics, such as small dataset issues, overfitting, and computational complexity. In order to guarantee the robustness and scalability of the model for clinical practice, the research incorporated data augmentation, external validation, and balanced dataset techniques. This is a significant breakthrough in the detection of hematological cancer, and this AI driven diagnostic system has a very high performance. It can potentially assist pathologists and oncologists in the early diagnosis and classification of leukemia and lymphoma with a single model and therefore lead to better patient outcomes and treatment rates. Future work should further extend the dataset to detect more diseases thus making it more robust, as well as the integration of this system into the healthcare as a real time diagnostic tool in clinical practice. This study introduces the application of AI in multiple blood cancer detection in hematological oncology and therefore represents a starting point for the development of fully automated, highly accurate, and easy to use cancer diagnostic systems. Declarations Ethical Statement This study does not contain any studies with human or animal subjects performed by any of the authors. Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Acknowledgments: The authors would like to thank the University of Salford for providing the funding for the article publication costs. Data Availability Statement The data that support the findings of this study as well as the analysis and model training are openly available in the repository “ALL-Detection” specifically in the notebook file named “lymphoma-leukemia-detection.ipynb” accessed via the link lymphoma-leukemia-detection References Kwee TC, Kwee RM, Nievelstein RAJ (2008) Imaging in staging of malignant lymphoma: A systematic review. Blood 111:1072–1082 Malard F, Mohty M (2020) Acute lymphoblastic leukaemia. Lancet 395:1146–1162 Coustan-Smith E, Mullighan CG, Onciu M et al (2011) New markers for minimal residual disease detection in acute lymphoblastic leukemia. Blood 117:6267–6276 Shalabi H, Angiolillo AL, Fry TJ et al (2020) Disease detection methodologies in relapsed B-cell acute lymphoblastic leukemia: Opportunities for improvement. Pediatr Blood Cancer 67:e28149 Wijetunga NA, Imber BS, Caravelli JF, Mikhaeel NG (2021) Radiomics in lymphoma: A systematic review. Br J Radiol 94:20210285 Prasad P, Anbarasi LJ (2024) Acute Lymphoblastic Leukemia Subtypes Detection using Vision Transformer Model. In: Proceedings of the 5th International Conference on Data Intelligence and Cognitive Informatics, ICDICI 2024. Institute of Electrical and Electronics Engineers Inc., 1413–1418 Albeeshi AA, Alshanbari HS (2023) Modeling of the Acute Lymphoblastic Leukemia Detection by Convolutional Neural Networks (CNNs). Curr Med Imaging Rev 19:835–844 Ibtekar T, Haque MR, Srizon AY (2023) Comparative Study Between ResNet And EfficientNet Family For Classification of Leukemia. In: 2023 26th International Conference on Computer and Information Technology (ICCIT). IEEE, 1–6 Rezayi S, Mohammadzadeh N, Bouraghi H, Saeedi S, Mohammadpour A (2021) Timely Diagnosis of Acute Lymphoblastic Leukemia Using Artificial Intelligence-Oriented Deep Learning Methods. Comput Intell Neurosci 2021:5478157 Abunadi I, Senan EM (2022) Multi-Method Diagnosis of Blood Microscopic Sample for Early Detection of Acute Lymphoblastic Leukemia Based on Deep Learning and Hybrid Techniques. Sens (Basel) 22:1629 Masilamani V, Devanesan S, AlSalhi MS, AlQahtany FS, Farhat KH (2020) Fluorescence spectral detection of acute lymphoblastic leukemia (ALL) and acute myeloid leukemia (AML): A novel photodiagnosis strategy. Photodiagnosis Photodyn Ther 29:101634 Khandekar R, Shastry P, Jaishankar S, Faust O, Sampathila N (2021) Automated blast cell detection for Acute Lymphoblastic Leukemia diagnosis. Biomed Signal Process Control 68:102690 Genovese A, Hosseini MS, Piuri V, Plataniotis KN, Scotti F (2021) Histopathological transfer learning for acute lymphoblastic leukemia detection. In: CIVEMSA 2021 - IEEE International Conference on Computational Intelligence and Virtual Environments for Measurement Systems and Applications, Proceedings. Institute of Electrical and Electronics Engineers Inc Nautiyal U, Bhatt A, Chauhan R, Rawat R, Gupta R (2024) A Modified Conventional Neural Network for Detecting and Classifying Leukemia. In: Proceedings of the 2nd International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics, ICIITCEE 2024. Institute of Electrical and Electronics Engineers Inc Albano D, Bruno A, Patti C, Micci G, Midiri M, Tarella C, Galia M (2019) Whole-body magnetic resonance imaging (WB-MRI) in lymphoma: State of the art. Hematol Oncol 38:12–21 Weisman AJ, Kieler MW, Perlman SB, Hutchings M, Jeraj R, Kostakoglu L, Bradshaw TJ (2020) Convolutional neural networks for automated pet/ct detection of diseased lymph node burden in patients with lymphoma. Radiol Artif Intell 2:e200016 Steinbuss G, Kloss M, Laleh G, Wagner M, Zgorzalski W, Bol K, Lapi E, Lengerke C, Tatsch T, Müller A (2021) Deep learning for the classification of non-hodgkin lymphoma on histopathological images. Cancers (Basel) 13:2419 Sheng B, Zhou M, Hu M, Li Q, Sun L, Wen Y (2020) A blood cell dataset for lymphoma classification using faster R-CNN. Biotechnol Biotechnol Equip 34:413–420 Kanavos T, Birbas E, Zanos TP (2025) A Systematic Review of the Applications of Deep Learning for the Interpretation of Positron Emission Tomography Images of Patients with Lymphoma. Cancers (Basel) 17:69 Miyoshi H, Maeshima AM, Nomoto J, Tng N, Tsuruoka S, Kono S, Nakai T, Fukuhara S, Munakata W, Maruyama D, Tobinai K, Kobayashi Y, Kober T (2020) Deep learning shows the capability of high-level computer-aided diagnosis in malignant lymphoma. Lab Invest 100:1300–1310 Yang YF, Zhao E, Shi Y, Zhang H, Yang YY (2024) Multicenter investigation of preoperative distinction between primary central nervous system lymphomas and glioblastomas through interpretable artificial intelligence models. Neuroradiology 66:1765–1774 Bai A, Si M, Xue P, Qu Y, Jiang Y (2024) Artificial intelligence performance in detecting lymphoma from medical imaging: a systematic review and meta-analysis. BMC Med Inf Decis Mak 24:14 Ghaderzadeh M, Aria M, Hosseini A, Asadi F, Bashash D, Abolghasemi H (2022) A fast and efficient CNN model for B-ALL diagnosis and its subtypes classification using peripheral blood smear images. Int J Intell Syst 37:5113–5133 Orlov NV, Chen WW, Eckley DM, Macura TJ, Shamir L, Jaffe ES, Goldberg IG (2010) Automatic classification of lymphoma images with transform-based global features. IEEE Trans Inf Technol Biomed 14:1003–1013 Kaggle Multi Cancer Dataset. Available at https://www.kaggle.com/datasets/obulisainaren/multi-cancer-dataset . Accessed 26 July 2024 He K, Zhang X, Ren S, Sun J: Deep Residual Learning for Image Recognition. In: IEEE Conference on Computer Vision and, Recognition P (2016) (CVPR). IEEE, 770–778, 2016 Wang MH, Chong KKL, Lin Z, Yu X, Pan Y (2023) An Explainable Artificial Intelligence-Based Robustness Optimization Approach for Age-Related Macular Degeneration Detection Based on Medical IOT Systems. Electron (Switzerland) 12:2697 Ullah MZ, Zheng Y, Song J, Aslam S, Xu C, Kiazolu GD, Wang L (2021) An Attention-Based Convolutional Neural Network for Acute Lymphoblastic Leukemia Classification. Appl Sci (Basel) 11:10662 Huang X, Jeon H, Liu J, Yao J, Wei M, Han W, Chen J, Sun L, Han J (2021) Deep-Learning Based Label-Free Classification of Activated and Inactivated Neutrophils for Rapid Immune State Monitoring. Sens (Basel) 21:512 Alim MS, Bappon SD, Sabuj SM, Islam MJ, Tarek M, Azam MS, Islam MM (2024) Integrating convolutional neural networks for microscopic image analysis in acute lymphoblastic leukemia classification: A deep learning approach for enhanced diagnostic precision. Syst Soft Comput 6:200121 Ryu D, Bak TH, Ahn D, Kang HY, Oh S, Min HS, Lee S, Lee J (2023) Deep Learning-Based Label-Free Hematology Analysis Framework Using Optical Diffraction Tomography. Heliyon 9:e18297 Reena MR, Ameer PM (2022) A content-based image retrieval system for the diagnosis of lymphoma using blood micrographs: An incorporation of deep learning with a traditional learning approach. Comput Biol Med 145:105463 Boldú L, Merino A, Acevedo A, Molina A, Rodellar J (2021) A deep learning model (ALNet) for the diagnosis of acute leukaemia lineage using peripheral blood cell images. Comput Methods Programs Biomed 202:105999 V KD SG (2024) M G, M KS, A MC, S R: Advanced Machine Learning Techniques for Accurate Detection and Diagnosis of Acute Lymphoblastic Leukemia (ALL). In: 2024 5th International Conference on Electronics and Sustainable Communication Systems (ICESC). IEEE, 1362–1366 Syrykh C, Abreu A, Amara N, Siegfried A, Maisongrosse V, Frenois FX, Martin L, Rossi C, Laurent C, Brousset P (2020) Accurate diagnosis of lymphoma on whole-slide histopathology images using deep learning. NPJ Digit Med 3:63 Zhou Z, Jain P, Lu Y, Macapinlac H, Wang ML, Son JB, Pagel MD, Xu G, Ma J (2021) Computer-aided detection of mantle cell lymphoma on 18F-FDG PET/CT using a deep learning convolutional neural network. Am J Nucl Med Mol Imaging 11:260–270 Achi HE, Belousova T, Chen L, Wahed A, Wang I, Hu Z, Kanaan Z, Rios A, Nguyen AND (2019) Automated Diagnosis of Lymphoma with Digital Pathology Images Using Deep Learning. Ann Clin Lab Sci 49:153–160 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8155029","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":547523923,"identity":"319f6fd9-ab1f-45a1-9c1e-0de50f485da0","order_by":0,"name":"Benedict Onochie Ibe","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/0lEQVRIiWNgGAWjYBACAxCR2MAgx8CQAGJa8MMEGSQIaDGGapGQbCBKC2MDSBexWswZmA8wPNxhk97PnmPA+KNGQoJf+vAGhh81DIkzG7BrsWxgS2BIPJOWO7PnjQEzzzEJCcm+tALGnmMMibNxOewAjwFDYtvh3A03cgyYGdgk6gzOAEV4gU6dh1/L/3T7GyCH/ZOQAGlh/EtYy4EEA4kcoOFtEC3MIFtwOcyymS3hQOKZZMMZZ54VHObtA/qlh63gsMwxCWNc3jdnbz748OcOO3n+9uSND398s5Hg52He+PBNjY3sjAM4rGFmYIBLITNwxcooGAWjYBSMAmIAAMmCVFH2w+AwAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0006-8745-0881","institution":"University Of Salford Manchester, UK","correspondingAuthor":true,"prefix":"","firstName":"Benedict","middleName":"Onochie","lastName":"Ibe","suffix":""},{"id":547523924,"identity":"2f5dcb16-1ccc-4565-a602-2a401179def3","order_by":1,"name":"Dagogo Godwin Orifama","email":"","orcid":"","institution":"University Of Salford Manchester, UK","correspondingAuthor":false,"prefix":"","firstName":"Dagogo","middleName":"Godwin","lastName":"Orifama","suffix":""},{"id":547523925,"identity":"d70db4f0-3728-4fb0-a191-36f3958db588","order_by":2,"name":"Ali Dan","email":"","orcid":"","institution":"University Of Salford Manchester, UK","correspondingAuthor":false,"prefix":"","firstName":"Ali","middleName":"","lastName":"Dan","suffix":""},{"id":547523926,"identity":"a66e1326-0aaa-4d2b-a9a5-30eafa4b0dd0","order_by":3,"name":"Ikechukwu Nwagbo Enumah","email":"","orcid":"","institution":"Edinburgh Napier University, Scotland","correspondingAuthor":false,"prefix":"","firstName":"Ikechukwu","middleName":"Nwagbo","lastName":"Enumah","suffix":""},{"id":547523927,"identity":"3b7d4390-0acc-4286-a969-4a93f38bf14b","order_by":4,"name":"Dominic Chinedu Ogbuagu","email":"","orcid":"","institution":"University Of Salford Manchester, UK","correspondingAuthor":false,"prefix":"","firstName":"Dominic","middleName":"Chinedu","lastName":"Ogbuagu","suffix":""},{"id":547523928,"identity":"b28ee0da-74ed-4f62-8d5c-c3e7e5f6ec56","order_by":5,"name":"Gbubemi Erics","email":"","orcid":"","institution":"University Of Salford Manchester, UK","correspondingAuthor":false,"prefix":"","firstName":"Gbubemi","middleName":"","lastName":"Erics","suffix":""}],"badges":[],"createdAt":"2025-11-19 12:07:33","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8155029/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8155029/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96587387,"identity":"c1df893d-c0b2-49e4-b79e-51480b396d71","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":782595,"visible":true,"origin":"","legend":"","description":"","filename":"DeepLearningModelsforAccurateLeukemiaLymphomaDetectionrevised.editedV2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/91e77d345cb97c2f435419e9.docx"},{"id":96587400,"identity":"cd11fdd5-8ee1-479b-89f4-f34105992bbf","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8155029.json","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/929e30ad78801c8334e5a874.json"},{"id":96587389,"identity":"808a8a54-6301-4e54-8153-e177819f2d2f","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":132606,"visible":true,"origin":"","legend":"","description":"","filename":"rs81550290enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/606618aa3de98d7d1a8274dc.xml"},{"id":96587390,"identity":"7781bfcc-ec3e-43a6-910f-705c61264d67","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/989abd6a49c047c2e6540b36.jpeg"},{"id":96605806,"identity":"ad5ff8bf-96d4-491f-a52e-11bcb245780e","added_by":"auto","created_at":"2025-11-24 09:24:04","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18670,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/df5814fa1565dc8c0a3d611a.png"},{"id":96605694,"identity":"3c5cfd26-f20d-403d-8537-3205a3593a11","added_by":"auto","created_at":"2025-11-24 09:23:49","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18214,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/7c76e04ff6b3199cda6561a7.png"},{"id":96587382,"identity":"c893d327-20cb-4545-bef1-5810978f8aef","added_by":"auto","created_at":"2025-11-24 05:19:24","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130499,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/0082f5322862536ddd93f53c.jpeg"},{"id":96587411,"identity":"28fd0be4-7c25-431f-a307-c7bef0532cc5","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11467,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/01cdf30e17834745da586e6e.png"},{"id":96605130,"identity":"cb44beb3-5f5d-402b-a5bc-3b79132c252c","added_by":"auto","created_at":"2025-11-24 09:18:53","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23604,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/403210033015386e80a2a41c.png"},{"id":96587396,"identity":"6ef868a7-f7aa-4343-9204-d1ff51dd1759","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20320,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/196b42246319210a4e5a8931.png"},{"id":96605352,"identity":"11e75a6a-bdfa-4e6e-88a0-972968115e2c","added_by":"auto","created_at":"2025-11-24 09:22:28","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10677,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/328e9e55fb76797aa8195920.png"},{"id":96587404,"identity":"283e9ffa-05d0-44fa-be6e-fd56791a13dc","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"jpeg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":145573,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage17.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/a36ce641bbf795d9b8fc48a3.jpeg"},{"id":96605018,"identity":"9059be1d-d0f2-4fda-b400-3e6f5a098fc0","added_by":"auto","created_at":"2025-11-24 09:17:29","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14483,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/62a74255790653a4c3ff3a90.png"},{"id":96587425,"identity":"23b7d9f9-60e0-4fb3-b70c-eaf248257e98","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"jpeg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":207495,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage19.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/a52b6132d1f6baae2c8d5830.jpeg"},{"id":96587422,"identity":"09e1a028-42a8-49b0-b4d2-2c44fcd48ad1","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7788,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/be32fec8fb59590200ab6361.jpeg"},{"id":96587423,"identity":"07e3df91-44bc-4db0-a5a7-8d4393def00b","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15252,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage20.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/c6e839c04ddab1b69d82c76b.png"},{"id":96587399,"identity":"305d43b9-5328-4b23-a323-644f3003f3e7","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"jpeg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":144318,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage21.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/64528576e04e506e68f9862b.jpeg"},{"id":96587428,"identity":"4c82f8f7-0a26-49ca-bea2-93b01e7f7d67","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"jpeg","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1644,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/c0f334c7bf4b0a54e1597ac3.jpeg"},{"id":96587401,"identity":"076525c2-8292-4928-b326-b5429e581188","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69032,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/53690087554d1f958feb0c45.png"},{"id":96587398,"identity":"ecc802aa-2197-4f63-a9f8-bb1b85003402","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":143269,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/a3b385381ac49ced1fd62084.png"},{"id":96587410,"identity":"58e1d5f0-0c4f-47c4-a486-838deb1bd2e2","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10369,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/47ca3bb66f43d2a01cf5f9b5.png"},{"id":96587408,"identity":"eaf27f15-f224-4949-8b23-89819e22925c","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"jpeg","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":145021,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/9ef5f2367837f69fb14da28b.jpeg"},{"id":96605426,"identity":"3a0c12ab-7864-4674-86c1-3db06a4e8515","added_by":"auto","created_at":"2025-11-24 09:22:51","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9904,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/3601ac00a31a008abc33f1d0.png"},{"id":96587415,"identity":"998ae4f1-4efe-46cf-9dd9-de0d8c7465b9","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25203,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/89fee4ce30ab14da5d91b773.png"},{"id":96587413,"identity":"68476892-b379-4465-84f2-8fe6293029da","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/0a5ec3d8760ebad4c37243ec.png"},{"id":96587436,"identity":"8e439889-242b-414f-9d65-2e6f54bb9592","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5723,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/79c140331b34d0f057d6e2a0.png"},{"id":96587393,"identity":"10a4c332-351b-4ab7-9b9e-d2a27c29844c","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2472,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/24a19d5cf595644941b76edd.png"},{"id":96587403,"identity":"8e2ed7fb-6e0b-46e8-818d-1d5c7a20b48e","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26096,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/2168020e9940e8e22fc117fb.png"},{"id":96587412,"identity":"e82acd10-3d19-4ae2-9c68-da9f11520f32","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2276,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/9ff4a8ae0d0677b822b71e47.png"},{"id":96587406,"identity":"c7bea09f-52b1-40ef-9aed-c4bf8efc51ab","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8687,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/0d3267e6f99c429456ae4558.png"},{"id":96587430,"identity":"c99e6baa-ca42-4f6c-b1cf-9d4d65821dd9","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5930,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/8678d7e0f229b40891e427d4.png"},{"id":96605747,"identity":"983f4457-1854-4c76-b63e-9273802d0b09","added_by":"auto","created_at":"2025-11-24 09:23:57","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2324,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/1da80ff65327c523dc027531.png"},{"id":96587420,"identity":"8ea079b4-b4ea-4cc4-b6ec-8ec6ea100ae3","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29667,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/3308aeeb38c33a9b3104cc00.png"},{"id":96605462,"identity":"85eda044-d4fb-4ea1-b5b5-c00782525933","added_by":"auto","created_at":"2025-11-24 09:23:06","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3022,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/d637d3dafa34c6ea0cb9b053.png"},{"id":96605942,"identity":"85c484a9-5b48-4be0-b90a-bb53bc151286","added_by":"auto","created_at":"2025-11-24 09:24:25","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41828,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage19.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/d8a785ee6f79610b5600b51f.png"},{"id":96587431,"identity":"8ed66a7a-7d07-4a3b-b1ed-e04815346fa5","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1159,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/8f154415691a2a1debadc2c7.png"},{"id":96605021,"identity":"c881b4fe-551c-4cb6-bcad-d3279aa1d7bd","added_by":"auto","created_at":"2025-11-24 09:17:30","extension":"png","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3200,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage20.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/a29550dcecbc9806d293f899.png"},{"id":96587433,"identity":"9243c3a2-cbb5-4da9-b68e-018964d0a230","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":31226,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage21.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/c64902ba218b3d2adf76cd77.png"},{"id":96587405,"identity":"6642b8b4-1d20-41a3-bb72-8b5463ea8e0e","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"png","order_by":39,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":960,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/aff22ba8a1054e0add2bbc99.png"},{"id":96587419,"identity":"dcca2217-58a5-428b-b762-9dcf60890e22","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":40,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14657,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/b5780ace7e211cf1b4185b4b.png"},{"id":96587435,"identity":"3da7c3dd-e473-480b-a4db-93e84adfdb77","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":41,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29319,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/636a67e6d927e9439f0dc4d3.png"},{"id":96587427,"identity":"93d1d21c-1649-4488-9dea-0b9cc8b2de64","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":42,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2439,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/f723f87764496f0a46e1a066.png"},{"id":96587414,"identity":"6af00cd4-011c-4766-9bda-4e0fa949f5fc","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"png","order_by":43,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":30979,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/1bd4e8a7068d1c289d58e8ee.png"},{"id":96587432,"identity":"c9c55534-edb2-4f2f-a9c8-9b02e014fd36","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":44,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2144,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/502dd2423253d84ce122de79.png"},{"id":96587438,"identity":"f5335f34-704e-48bc-9aff-95918145ed9b","added_by":"auto","created_at":"2025-11-24 05:19:27","extension":"png","order_by":45,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9495,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/741880c25675cec47e7f607a.png"},{"id":96605794,"identity":"ec65dc41-a80b-4b9a-be05-5f0fe17b305c","added_by":"auto","created_at":"2025-11-24 09:24:03","extension":"xml","order_by":46,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":126418,"visible":true,"origin":"","legend":"","description":"","filename":"rs81550290structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/e966433413f77703a36053f8.xml"},{"id":96605274,"identity":"7643a03d-f82a-410b-82b1-2d041021461b","added_by":"auto","created_at":"2025-11-24 09:21:57","extension":"html","order_by":47,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":144748,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/d36cb2758a12491232a0d450.html"},{"id":96587386,"identity":"946c9ed8-14b4-4577-965b-396465eaf73e","added_by":"auto","created_at":"2025-11-24 05:19:24","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":71008,"visible":true,"origin":"","legend":"\u003cp\u003eBalancing the Data: (a) Imbalanced Dataset and (b) Balanced Dataset After Oversampling\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/1ecad7e5c4d1ea68c59ac4d0.jpg"},{"id":96587417,"identity":"c89981ba-d12a-436e-9503-aaa97b2bd6e7","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":132605,"visible":true,"origin":"","legend":"\u003cp\u003eResearch Methodology Diagram, which details the full implementation or process flow of the research\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/2bf63431294b8a9e7cfc563f.jpg"},{"id":96587416,"identity":"a8b2ea26-0fdb-4dda-867a-a1cab600a1cb","added_by":"auto","created_at":"2025-11-24 05:19:26","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":78300,"visible":true,"origin":"","legend":"\u003cp\u003eConvNeXt Performance: (a) ConvNeXt Loss and Accuracy Curve (b) ConvNeXt Confusion Matrix\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/49ea45d57a6f3efb5cde9eb8.jpg"},{"id":96587383,"identity":"2979515a-1621-43d1-98c8-ca38dcd988a9","added_by":"auto","created_at":"2025-11-24 05:19:24","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":70182,"visible":true,"origin":"","legend":"\u003cp\u003eSWIN Performance: (a) SWIN Loss and Accuracy Curve (b) SWIN Confusion Matrix\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/6ae2ad0785dcb2258c82a3a9.jpg"},{"id":96604770,"identity":"880e6c13-0a7b-4d5a-9968-0256a11df9b4","added_by":"auto","created_at":"2025-11-24 09:14:52","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":75408,"visible":true,"origin":"","legend":"\u003cp\u003eResNet101 Performance: (a) ResNet101 Loss and Accuracy Curve (b) ResNet101 Confusion Matrix\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/d897d1760d46dabf0c82d76f.jpg"},{"id":96605316,"identity":"351ee975-57f6-4a54-8519-66f749bbe044","added_by":"auto","created_at":"2025-11-24 09:22:18","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":68229,"visible":true,"origin":"","legend":"\u003cp\u003eVGG16 Performance: (a) VGG16 Loss and Accuracy Curve (b) VGG16 Confusion Matrix\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/b667a70d0d6a1365a0c5c614.jpg"},{"id":96605012,"identity":"e663c30e-32c0-470e-8217-a4e2349b8596","added_by":"auto","created_at":"2025-11-24 09:17:22","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":78811,"visible":true,"origin":"","legend":"\u003cp\u003eCustom ALL Performance: (a) Custom ALL Loss and Accuracy Curve (b) Custom ALL Confusion Matrix\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/277b87e40c1a2a0d5f048cb4.jpg"},{"id":96604917,"identity":"91c3740c-43f0-4071-bd7f-e7689a725e40","added_by":"auto","created_at":"2025-11-24 09:16:02","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":114686,"visible":true,"origin":"","legend":"\u003cp\u003eConvNext (ALL+Lymphoma) Performance: (a) ConvNext (ALL+Lymphoma) Loss and Accuracy Curve (b) ConvNext (ALL+Lymphoma) Confusion Matrix\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/7bebd21b3742996bdabf98e6.jpg"},{"id":96587392,"identity":"65965303-e345-43c9-ac70-97c829ddc86c","added_by":"auto","created_at":"2025-11-24 05:19:25","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":80362,"visible":true,"origin":"","legend":"\u003cp\u003eConvNeXt (ALL+Lymphoma) Performance: (a) ConvNeXt(ALL+Lymphoma) Test on the External Lymphoma Dataset, (b) ConvNeXt(ALL+Lymphoma) Test on the External ALL Dataset\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/144244dda49b5046b3ef082b.jpg"},{"id":96708090,"identity":"4ee90996-441c-41f1-ba64-380e79992fcd","added_by":"auto","created_at":"2025-11-25 09:56:21","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1980345,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8155029/v1/99bc4658-243f-4acc-a643-4c5bc392fe18.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eDeep Learning Models for Accurate Leukemia \u0026amp; Lymphoma Detection\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eHematologic malignancies like Acute Lymphoblastic Leukemia (ALL) and lymphoma have been diagnosed and detected with better methods over the last century. Initially, the diagnosis was made based on clinical findings and routine histological examination. The progression of imaging modalities from simple X-ray and lymphography to sophisticated CT, FDG-PET, and PET/CT has significantly enhanced the sensitivity and specificity of lymphoma staging and assessment of treatment outcomes [1]. Additionally, in the case of ALL, the combination of molecular diagnostics, genetic tests, and MRD detection using flow cytometry and PCR has altered the approach to risk categorization and disease management [2] [3]. These advancements have not only increased survival rates but also facilitated the development of personalized treatment plans, particularly with the aid of immunotherapies targeted at specific genetic mutations.\u003c/p\u003e\u003cp\u003eNevertheless, there are still issues regarding the improvement of these diagnostic methods for clinical practice. In the diagnosis of lymphoma, PET/CT fusion is superior to other modalities; however, variability in methods and high costs have limited its widespread use. Likewise, the use of MRD markers in ALL has been found to help detect early relapse; however, the role of these biomarkers in various patient populations, including adults, is currently under further investigation [4]. New technologies, such as radiomics and deep learning in imaging, as well as next-generation sequencing for genetic profiling, are showing potential in enhancing diagnostic accuracy and personalized care in cancer management [5]. Nevertheless, there is a need to establish the validity of these new technologies and to eliminate the existing differences in diagnostic care between different healthcare settings.\u003c/p\u003e\u003cp\u003eEmerging architectures such as Vision Transformers (ViTs) and hybrid models like ConvNext offer promising solutions. ViTs utilize self-attention mechanisms to capture intricate relationships across different image regions, outperforming CNNs in haematological image classification tasks. For instance, ViTs achieved an accuracy of 98% [6], compared to 92.27% at a learning rate of 0.003 [7] for CNNs, highlighting their potential in enhancing diagnostic accuracy for ALL. Similarly, ConvNeXT models combine convolutional layers with attention mechanisms, optimizing feature extraction and performance.\u003c/p\u003e\u003cp\u003eDespite these advancements, challenges remain in deploying these models in clinical settings, particularly in resource-constrained environments. Lightweight architectures, such as MobileNet and SqueezeNet, provide a balance between accuracy and computational efficiency, making them suitable for rapid diagnostics. However, their performance is generally lower than that of more complex models. Pre-trained transfer models, including ResNet101, VGG16, and EfficientNet [8], have demonstrated high accuracy due to their rich feature representations. However, the need for extensive computational resources and large labelled datasets limits their widespread clinical application.\u003c/p\u003e\u003cp\u003eTherefore, there is a critical need for a diagnostic model that integrates the strengths of advanced deep learning architectures, such as the accuracy of ViTs, the hybrid efficiency of ConvNeXT, and the computational economy of lightweight models, while minimizing their limitations. This study addresses this gap by developing an optimized diagnostic framework specifically tailored for the detection and classification of ALL, ultimately aiming to enhance diagnostic accuracy, efficiency, and accessibility in diverse clinical settings.\u003c/p\u003e\u003cp\u003e\u003cb\u003e1.1. Objectives of the Study\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eTo develop a unified deep learning framework for the dual detection of ALL and Lymphoma\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eTo validate the proposed ConvNeXt-based framework through an ablation study against other prominent architectures.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eTo assess the framework's generalizability on independent, unseen external datasets.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eBuild one single Model to detect or predict the different stages of Acute Lymphoblastic Leukemia (Benign, Early, Pre, and Pro), and the different classes of Lymphoma (chronic lymphocytic leukemia, follicular lymphoma, and mantle cell lymphoma), thereby creating one Model to detect two types of ALL and Lymphoma.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"2. Literature Review","content":"\u003ch2\u003e2.1. Advancements in Acute Lymphoblastic Leukemia (ALL) Detection\u003c/h2\u003e\u003cp\u003eAcute Lymphoblastic Leukemia (ALL) is a type of cancer that affects the lymphoid cells. Conventional methods of diagnosis are based on blood smear analysis and immunophenotyping, which are time-consuming and dependent on the availability of pathologists. However, with the development of AI and deep learning, several innovative strategies have been proposed to enhance the diagnostic accuracy and speed of ALL.\u003c/p\u003e\u003cp\u003eDeep learning models, especially Convolutional Neural Networks (CNNs), have been applied to the detection of ALL from microscopic images. A study using ResNet-50, VGG-16, and custom convolutional networks achieved the highest validation accuracy of 84.62%, highlighting the potential of AI in automated leukemia diagnosis [9]. Other models have attempted hybrid approaches, such as combining CNN with SVM, which achieves nearly 100% accuracy in some cases [10]. In addition, fluorescence spectral methods for non-invasive ALL diagnosis have 88% sensitivity and 80% specificity, offering a new approach to distinguish leukemia from standard samples [11].\u003c/p\u003e\u003cp\u003eAutomated detection techniques using YOLOv4 have also been introduced, achieving a mean average precision (mAP) of 98.7%, proving the feasibility of object detection models for blast cell identification [12]. Transfer learning techniques have been applied to histopathological images, resulting in a significant improvement in classification accuracy [13]. However, challenges remain in standardizing datasets and ensuring the generalizability of models across diverse populations.\u003c/p\u003e\u003cp\u003e[14] Investigates the application of deep learning models for leukemia classification with the aid of AI, specifically DenseNet201, EfficientNet B2, and ConvNeXT Small. Of the three models, the ConvNeXT model had the highest accuracy of 92.88%, sensitivity of 83.89%, specificity of 97.07%, and F1-score of 88.22%. These results indicate that the ConvNeXT model is more effective at identifying leukemia cases and distinguishing between subtypes compared to the other two models. Hence, the proposed model holds potential for assisting clinicians in making early and accurate diagnoses of leukemia, which is crucial for managing this deadly disease.\u003c/p\u003e\u003cp\u003eResearch [15] has been conducted on various deep learning models for leukemia classification, including ConvNeXt, DenseNet201, EfficientNet B2, and ensemble models. Accuracy was highest for ConvNeXt at 92.8%, making it the best architecture for detecting leukemia. The study also highlights the significance of early identification and the feasibility of AI-based solutions for medical diagnosis.\u003c/p\u003e\u003ch2\u003e2.2. Advancements in Lymphoma Detection\u003c/h2\u003e\u003cp\u003eLymphoma, a cancer of the lymphatic system, has also been diagnosed more accurately with the help of AI-driven approaches. The traditional methods include whole-body PET/CT imaging and histopathological examination, but these are often time-consuming and prone to interobserver variability. AI has been integrated with lymphoma detection and deep learning-based automated PET/CT analysis, which achieved an 85% accurate positive rate (TPR) using CNNs [16].\u003c/p\u003e\u003cp\u003eHistopathological image analysis using CNNs has also been effective in lymphoma subtype classification, with EfficientNet-based models reaching 95.56% accuracy in distinguishing different NHL subtypes [17]. Faster R-CNN models have been trained on blood cell datasets, yielding over 96% detection rates for lymphoma cells [18]. Moreover, a systematic review on deep learning in PET image interpretation found AI models achieving an AUC of 0.963 in lymphoma detection [19].\u003c/p\u003e\u003cp\u003eHowever, there are still issues with the size of the datasets, the absence of external validation, and the necessity to translate the models for real clinical use. Although AI models are accurate in diagnosis, their integration into regular clinical practice and comparison with human pathologists has not been fully understood [20].\u003c/p\u003e\u003cp\u003eThis study [21] aims to compare the preoperative differentiation between Primary Central Nervous System Lymphoma (PCNSL) and Glioblastoma (GBM) using deep learning and radiomics-based AI models. Three models were evaluated: The Radiomics model (AUC = 0.86), the ConvNeXt model (AUC = 0.89), and the MobileVIT Model (AUC = 0.91). The Max-Fusion Model, which combines radiomics and deep learning, achieved the best performance with an AUC of 0.92. This fusion-based approach significantly enhances tumor classification accuracy and may thus aid clinicians in making non-invasive decisions. The model that performed the best was the Max-Fusion model, which combines radiomics and deep learning, with an AUC of 0.92. This fusion-based approach significantly improves tumor classification accuracy and, in turn, may be helpful to clinicians in making non-invasive decisions.\u003c/p\u003e\u003ch2\u003e2.3. Problems\u003c/h2\u003e\u003cp\u003eHowever, several problems persist. Many of the models are trained on the institution-specific small datasets, which are not generalizable. Furthermore, there is no standardized evaluation framework for such tasks, making it difficult to compare different approaches. The lack of external validation is also a concern that raises questions about the real-world applicability of these models. The aforementioned challenges include overfitting and the need for high computational resources, both of which are discouraging for widespread clinical adoption. The problems or challenges are explained in more detail in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e below:\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eProblems identified in previous research about the implementation of Artificial Intelligence and Deep learning\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eProblem/Gaps Identified\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eValidation from Research Sources\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLack of Large-Scale Datasets\u003c/b\u003e - Many AI models are trained on small, institution-specific datasets, which limits generalizability.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSource: A systematic review found that many studies lacked large datasets and external validation, which limited their clinical application [22]. Also, the PET/CT lymphoma detection study [16] highlighted that training CNNs on limited datasets affects performance\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLimited External Validation\u003c/b\u003e - Many models are trained and tested on the same dataset, without independent validation.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSource: The review of AI in PET imaging highlighted the lack of external validation in lymphoma detection models [19]. Additionally, the histopathological AI model study mentioned that it had limited testing on diverse datasets [17].\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eOverfitting in Deep Learning Models -\u003c/b\u003e Some AI models achieve near-perfect accuracy on training data but may not generalize to real-world cases.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSource: The YOLOv4 leukemia detection study mentioned that deep learning models require careful tuning to avoid overfitting [12]. Also, the ALL deep learning comparison highlighted concerns about model robustness [9].\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eComputational Resource Requirements -\u003c/b\u003e Many deep learning models require expensive GPUs and substantial computing power.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSource: The histopathological AI study mentioned computational requirements as a challenge for real-world adoption [17]. Additionally, studies utilizing deep CNNs have emphasized the need for specialized hardware [9].\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003cb\u003e2.4. Proposed Solution For The Research\u003c/b\u003e\u003c/p\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eGather data from several sources to build a big and diverse dataset to increase the model's generalization for ALL and lymphoma detection.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eEnhance model generalization with external validation, which ensures that the models are trained on one dataset and validated on completely different datasets that have not been seen before for real-world applications.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eData augmentation, cross-validation, and regularization are employed to enhance the model's robustness and prevent overfitting.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAblation or comparative studies using Transformers (SWIN), ConvNeXT, and CNN-based architectures for high accuracy, efficiency, and computational optimization.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDesign a single AI system that can identify various stages of ALL (Benign, Early, Pre, Pro) and types of lymphoma (CLL, FL, MCL).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cp\u003eTo achieve these objectives, we propose a framework centered on the ConvNeXt architecture, whose hybrid design is uniquely suited to overcome the feature extraction challenges in hematological imaging.\u003c/p\u003e"},{"header":"3. Research Methodology","content":"\u003cp\u003eThe initial phase of this study focused on detecting Acute Lymphoblastic Leukaemia (ALL) and comparing the effectiveness of different deep learning models on an ALL dataset. The images were taken from Kaggle [23] and included four stages of ALL: Benign, Early, Pre, and Pro. The images were available in two forms: The original and the segmented, to see how the feature extraction differs. The images were imported into a pandas DataFrame, named by class labels and image type, and EDAs were conducted on the data. To identify imbalances, a class distribution analysis was conducted, and sample images from each class were displayed to gain an understanding of the dataset's structure.\u003c/p\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Data Preprocessing\u003c/h2\u003e\u003cp\u003eSeveral preprocessing steps were undertaken to standardize the input data:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003ePath Setup and Data Loading: The dataset's structure was also explored, and a DataFrame was generated to store the image paths, class labels, and image types.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eLabel Encoding: Both ALL and Lymphoma class labels were numerically valued in the following:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eALL Classes: Benign (0), Early (1), Pre (2), Pro (3)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eLymphoma Classes: CLL (4), FL (5), MCL (6)\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eData Splitting: The dataset was divided into training (80%), validation (10%), and test (10%) sets, created using stratified sampling to maintain class balance.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eImage Preprocessing Function: A specific function was written to import and resize the images to 224x224 pixels for uniformity with other networks. This function also undecoded and denormalized the images to match the chosen architectures.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eShuffling and Batching: The training, validation, and test sets were randomly shuffled to introduce some element of randomness and to mitigate any potential bias during training. An interval of 32 was chosen for the batch size to optimize memory utilization and model training time.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eClass Balancing: To check for any class imbalance, the dataset was grouped by class labels, and the distribution of images per class was visualized using bar charts. This step was crucial to ensure that each class had a representative number of samples, which informed the class balancing techniques employed later in the preprocessing pipeline, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(a) below. The dataset class imbalance was therefore addressed using oversampling, which was applied through resampling techniques to ensure uniform representation across classes, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(b).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Data Augmentation\u003c/h2\u003e\u003cp\u003eTo enhance the robustness of the models, data augmentation strategies were used during the training process:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRandom Flipping\u003c/b\u003e: To create different views, images were flipped horizontally.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRandom Rotation\u003c/b\u003e: A small amount of random degree rotation (up to 20%) was applied to the images to account for variation in orientation.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eNormalization\u003c/b\u003e: All pixel values were normalized to the range of 0 to 1 to guarantee stable training.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eAfter the preprocessing steps, we conducted an initial comparative analysis of multiple deep learning models for ALL detection, comparing the performance of the Swin model, ResNet101, VGG16, a Custom CNN model, and the ConvNeXt model. It was observed that ConvNeXt achieved the highest accuracy. However, to extend the applicability of the model beyond ALL detection, the dataset was expanded to include Lymphoma. This inclusion aimed to create a comprehensive haematological model using the ConvNeXt model due to the fact that it performed best in the initial comparative analysis, capable of detecting both ALL and Lymphoma, thereby improving diagnostic utility.\u003c/p\u003e\u003cp\u003eThe Lymphoma dataset was downloaded from Kaggle [24] using the Kaggle API, specifically from the Malignant Lymphoma Classification dataset. The dataset consisted of images labelled under three primary lymphoma classes: Chronic Lymphocytic Leukaemia (CLL), Follicular Lymphoma (FL), and Mantle Cell Lymphoma (MCL). The Lymphoma images were preprocessed similarly to the ALL dataset, and TIFF images were converted to JPG format for consistency.\u003c/p\u003e\u003cp\u003eTo ensure model generalization and prevent bias, the ALL and Lymphoma datasets were merged into a single dataset, resulting in a combined dataset containing images from seven classes (four ALL classes and three Lymphoma classes). The dataset was then checked for class imbalance, and to address any disparities, oversampling techniques were applied using resampling to balance the classes before training.\u003c/p\u003e\u003cp\u003eTo ensure consistency in data formats, the Lymphoma dataset\u0026rsquo;s TIFF images were converted to JPG. The ALL and Lymphoma datasets were then merged into a single dataset, resulting in a seven-class classification problem (four ALL classes\u0026thinsp;+\u0026thinsp;three Lymphoma classes).\u003c/p\u003e\u003cp\u003eTo further validate the robustness of the model, an additional Lymphoma dataset from Kaggle [25] and an additional ALL dataset from Roboflow in Pascal VOC format were downloaded to serve as additional external test datasets. These datasets were used to test the model on unseen data.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Model Selection and Proposed Architecture\u003c/h2\u003e\u003cp\u003eFollowing dataset preparation and preprocessing, we examined a few of the best contemporary models to establish the best architecture for our intended diagnostic system. Based on this analysis, the ConvNeXt model would be our proposed method due to its novel hybrid structure. This selection was benchmarked and validated by other models (Swin Transformer, ResNet101, VGG16, and a baseline CNN) as part of an ablation study.\u003c/p\u003e\u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\u003ch2\u003e3.3.1. ConvNeXt Model\u003c/h2\u003e\u003cp\u003eConvNeXt is a recent convolutional model designed with a transformer-like architecture to enhance the efficiency of convolutional operations. It introduces depthwise convolutions, LayerNorm, and GELU activation, which enhance feature extraction. ConvNext is a modern convolutional architecture trained with pre-trained weights, where ConvNext was fine-tuned for the task of imaging leukemia stages and assigning them the correct class, based on both local and global features. The conventional convolution operation is given by (1) below:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:y\\:=\\:x\\:+\\:f\\left(x\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere f(x) is made up of:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eDepthwise Conv (kernel size 7\u0026times;7)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eLayer Normalization\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePointwise Conv (1\u0026times;1) expanding channels by 4\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eGELU activation\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePointwise Conv (1\u0026times;1) reducing channels.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eLayer Scale (learned per-channel scaling)\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eAnd then fully expressed mathematically as (2):\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:f\\left(x\\right)=\\:{\\text{P}\\text{W}\\text{C}\\text{o}\\text{n}\\text{v}}_{\\text{d}\\text{o}\\text{w}\\text{n}}\\left(GELU\\left({PWConv}_{up}\\left(LN\\left(DWConv\\left(x\\right)\\right)\\right)\\right)\\right)⊚\\:\\lambda\\:$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\lambda\\:\\)\u003c/span\u003e\u003c/span\u003e is a learnable parameter for layer scaling.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\u003ch2\u003e3.3.2. SWIN Transformer Model\u003c/h2\u003e\u003cp\u003eThe transformer-based architecture used in this paper was initialized with pretrained weights and fine-tuned on the ALL dataset. The model capacity was enhanced through attention mechanisms to leverage the ability to model long-range dependencies for the classification of leukemia stages. The Swin Transformer is designed with a hierarchical architecture, where an image is treated as a sequence of non-overlapping windows to model both local and global dependencies powerfully. Such shifting window mechanisms are effective in capturing fine details from small regions while providing a broader perception of the whole image. While its multi-head self-attention inside the window enabled robust feature extraction, it likely endowed the model with high accuracy in the classification of all stages. Its mathematical representation involves:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePatch Partitioning\u003c/b\u003e: The input image is divided into non-overlapping patches as seen in (3).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:\\text{P}\\text{a}\\text{t}\\text{c}\\text{h}\\:\\text{s}\\text{i}\\text{z}\\text{e}\\left(\\:=\\:P\\right),\\hspace{1em}\\text{w}\\text{h}\\text{e}\\text{r}\\text{e}\\:P\\:\\text{i}\\text{s}\\:\\text{t}\\text{h}\\text{e}\\:\\text{p}\\text{a}\\text{t}\\text{c}\\text{h}\\:\\text{d}\\text{i}\\text{m}\\text{e}\\text{n}\\text{s}\\text{i}\\text{o}\\text{n}.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eLinear Embedding\u003c/b\u003e: Each patch is linearly projected into a feature vector, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{W}_{p}\\)\u003c/span\u003e\u003c/span\u003e is the learnable weight, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{p}\\)\u003c/span\u003e\u003c/span\u003e is the Input patch, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{and\\:b}_{p}\\)\u003c/span\u003e\u003c/span\u003e is the learnable bias as seen in (4).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:{z}_{p}={W}_{p}\\cdot\\:{x}_{p}+{b}_{p}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eShifted Window Attention\u003c/b\u003e: Self-attention is computed within local windows, and windows are shifted between layers to enable cross-window connections, and is represented in (5) below:\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:\\text{Attention}\\left(\\varvec{Q},\\:\\varvec{K},\\:\\varvec{V}\\right)=\\text{Softmax}\\left(\\frac{\\varvec{Q}{\\varvec{K}}^{\\varvec{T}}}{\\sqrt{{d}_{k}}}\\right)\\varvec{V}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere Q, K, V are the Query, Key, and Value matrices, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{d}_{k}\\)\u003c/span\u003e\u003c/span\u003e is the Scaling factor (dimension of keys)\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\u003ch2\u003e3.3.3. ResNet101 Model\u003c/h2\u003e\u003cp\u003eResNet101 is a deep convolutional neural network architecture designed by Microsoft Research in 2015 [26]. It is noted for having a very high number of layers (101 in total) and yet preventing the vanishing gradient problem through the use of innovative residual connections. The main idea of ResNet101 is the residual block, where instead of learning unreferenced functions, each layer learns residual functions that depend on the layer inputs. These \"skip connections\" enable the training of significantly deeper networks than previously possible. The idea was built on the success of ResNet, which won several visual recognition challenges in 2015. The architecture comprises a single convolutional layer followed by 33 bottleneck blocks (each containing three convolutional layers), organized into four stages of increasing feature map depth (64, 128, 256, and 512 channels). The bottleneck design is implemented to reduce and then restore dimensions using 1\u0026times;1 convolutions, with a 3\u0026times;3 convolution in between to preserve dimensions while efficiently managing computational resources.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003e3.3.4. VGG16\u003c/h2\u003e\u003cp\u003eVGG16 is a convolutional neural network architecture developed by the Visual Geometry Group at Oxford University in 2014 [27] and characterized by its simplicity and depth, specifically comprising 16 weight layers (13 convolutional layers and three fully connected layers). The architecture is characterized by a consistent pattern of 3x3 convolutional layers with ReLU activation functions, followed by max-pooling layers for down-sampling. The number of filters is increased progressively (from 64 in early layers to 512 in deeper layers) while the spatial dimensions are reduced through pooling. One of the design philosophies that proved to be highly effective was the use of multiple small filters rather than fewer large ones.\u003c/p\u003e\u003cp\u003eTransfer learning was applied using ResNet-101 and VGG-16, which were pre-trained on the ImageNet dataset. These models were modified by replacing the fully connected layers with custom classifiers tailored to the ALL dataset. The initial layers were frozen to retain general image features, and the later layers were fine-tuned for our task.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section3\"\u003e\u003ch2\u003e3.3.5. Custom CNN Model\u003c/h2\u003e\u003cp\u003eA custom CNN architecture was built with several convolutional layers followed by max-pooling and fully connected layers. This model served as a baseline to compare the effectiveness of transfer learning and transformer-based models.\u003c/p\u003e\u003cp\u003eAfter evaluating these models on all detection tasks, the ConvNeXt model emerged as the best-performing architecture, achieving the highest accuracy across all tasks. Given this, the study was extended to detect both ALL and Lymphoma by training the ConvNeXt model on the expanded dataset, which is a combination of the ALL dataset and the Lymphoma dataset.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e3.3.6. Evaluation Metrics\u003c/h2\u003e\u003cp\u003e\u003cem\u003eAccuracy\u003c/em\u003e: Accuracy is defined as the percentage of the models' predictions (including both true positives and true negatives) that match the total number of cases. It is stated as: \"What percentage of all predictions did the model get right?\" However, accuracy can be misleading on imbalanced datasets where one class is significantly outnumbered by the other. Accuracy is represented mathematically as (6) below:\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{c}\\text{c}\\text{u}\\text{r}\\text{a}\\text{c}\\text{y}=\\:\\frac{\\text{T}\\text{P}+\\text{T}\\text{N}}{\\text{T}\\text{P}+\\text{T}\\text{N}+\\text{F}\\text{P}+\\text{F}\\text{N}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003ePrecision\u003c/em\u003e: Precision is the percentage of correct identifications. It is stated as: \"When the model outputs the positive class, what is the chance of being accurate?\" Precision is stated as the fraction of the true positives over the sum of the true and false positives. It is essential in cases where false positives are expensive, and is expressed as (7) below:\u003cdiv id=\"Equ7\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$$\\:\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}=\\:\\frac{TP}{TP\\:+\\:FP}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003eRecall\u003c/em\u003e: Recall (also known as sensitivity) measures the proportion of actual positives that are correctly identified. It answers: \"What percentage of actual positive cases did the model identify?\" Recall is calculated as true positives divided by the sum of true positives and false negatives. It is critical in scenarios where missing a positive case has serious consequences, such as disease detection. Recall is expressed mathematically by (8):\u003cdiv id=\"Equ8\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ8\" name=\"EquationSource\"\u003e\n$$\\:Recall\\:=\\:\\frac{TP}{TP\\:+\\:FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e8\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eF1 Score\u003c/strong\u003e\u003cp\u003eThe F1 score is the harmonic mean between precision and recall, which is a single value that considers both of them. It is on a scale of 0 to 1 and is especially useful when the goal is to find a tradeoff between precision and recall or when the classes are imbalanced. The F1 score is calculated as (9)\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv id=\"Equ9\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ9\" name=\"EquationSource\"\u003e\n$$\\:\\text{F}1-\\text{S}\\text{c}\\text{o}\\text{r}\\text{e}\\:=\\:2\\:\\times\\:\\:\\frac{\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}\\times\\:\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}{\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}+\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e9\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere TP is the True Positives, TN is the True Negatives, FP is the False Positives, And FN is the False Negatives\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\u003ch2\u003e3.3.7. Training ConvNeXt for ALL and Lymphoma Detection\u003c/h2\u003e\u003cp\u003eTo train the final model on the combined dataset, the following configurations were used:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eModel Architecture\u003c/b\u003e: ConvNeXt, initialized with pre-trained weights, was fine-tuned for multi-class classification across the seven ALL and Lymphoma classes.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eLoss Function\u003c/b\u003e: Categorical cross-entropy loss was applied.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eOptimizer\u003c/b\u003e: Adam optimizer (lr\u0026thinsp;=\u0026thinsp;0.0001) with a scheduler to reduce learning rate on plateau.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eTraining Configuration\u003c/b\u003e: The model was trained for \u003cem\u003e10 epochs\u003c/em\u003e with early stopping based on validation loss.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eEvaluation Metrics\u003c/b\u003e: Accuracy, precision, recall, and F1-score were used for assessment, alongside confusion matrices to visualize classification performance.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section3\"\u003e\u003ch2\u003e3.3.8. External Dataset Evaluation\u003c/h2\u003e\u003cp\u003eAfter training the comprehensive ConvNeXt model, its efficacy was tested on external datasets that were not part of the training process. To validate the model's robustness, two additional datasets were downloaded:\u003c/p\u003e\u003cp\u003e1. \u003cb\u003eA different Lymphoma dataset from Kaggle\u003c/b\u003e - This dataset was downloaded using the Kaggle API and was used to test how well the model generalizes to unseen Lymphoma cases.\u003c/p\u003e\u003cp\u003e2. \u003cb\u003eA different ALL dataset from Roboflow\u003c/b\u003e - This dataset served as a secondary validation set for assessing the model's performance on new ALL samples.\u003c/p\u003e\u003cp\u003eThe trained ConvNeXt model was evaluated on both external datasets, and accuracy metrics were generated to determine its performance. This step ensured that the model was not overfitting to the original dataset and was capable of detecting both ALL and Lymphoma in diverse datasets.\u003c/p\u003e\u003cp\u003eBy expanding the dataset and retraining the model, this research significantly enhances hematological diagnostics, enabling broader leukaemia and lymphoma detection using deep learning. The methodology outlined ensures full reproducibility for further research in leukemia and lymphoma classification. A schematic of the complete research methodology is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"4. RESULTS AND DISCUSSION","content":"\u003cp\u003eThe test results of the models revealed that the ConvNeXt model performed the best, achieving an impressive 99.69% accuracy, followed closely by the Swin model with an accuracy of 99.39%. Both models demonstrated superior performance in differentiating between the stages of ALL, likely due to their modern architecture design, which efficiently captured both local and global features within the images.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eConvNeXt Model (99.69%)\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eThis model performed exceptionally well due to its ability to capture spatial hierarchies and visual features efficiently. ConvNeXt's modern convolutional structure is designed to mimic the strengths of both convolutional and transformer models, allowing it to achieve state-of-the-art performance. The training accuracy and loss curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a). At the same time, the ConvNeXt model demonstrated high precision, with only a single misclassification of a 'Benign' sample as 'Pre' (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb)\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eSwin Model (99.39%)\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eTraining the SWIN transformer model obtained a very high accuracy of 99.39% which shows its efficacy in perfectly classifying ALL classes, which is proof that the balance of Swin is done between the tradeoff of accuracy and computational efficiency, more so for this medical image task. The training accuracy and loss curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(a). In contrast, the confusion matrix for the Swin vision transformer model, which showed it only misclassified the 'Benign' label as 'Early', is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e (b) below\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eResNet101 (98.47%)\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eAs expected, ResNet101, a deep residual network, also performed well, given its depth and ability to learn hierarchical features. However, its accuracy was slightly lower than ConvNeXt and Swin model, possibly due to overfitting on certain leukemia stages. The confusion matrix for the ResNet101 model, seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e (b), showed it misclassified the 'Early' label as 'Benign' the most (2) due to the very minimal difference between the two stages, while the training accuracy and loss can be seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e (a)\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eVGG16 (96.32%)\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eWhile VGG16 is known for its strong feature extraction capabilities, its performance lagged behind more modern architectures like ConvNeXt and ResNet101. This may be due to the simplicity of the VGG architecture, which does not incorporate the advanced residual or attention mechanisms found in newer models. The confusion matrix for the VGG16 model, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e(b), indicated that it misclassified the 'Benign' label as 'Early' the most (9). At the same time, the training accuracy and loss for the VGG16 are seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e (a).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eCustom ALL Model (90.80%)\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eThe combination model, which integrates features from multiple architectures, performed reasonably well but did not match the precision of individual models, such as ConvNeXt or Swin model. This lower accuracy may reflect the challenge of combining features from multiple models without over-complicating the learning process. The training accuracy and loss of the Custom ALL model is seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e (a), while the confusion matrix for the ResNet101 model, which showed it misclassified the 'Early' label as 'Pre' and 'Early' as 'Benign' the most (6) and is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e (b) below.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003e4.1. Confusion Matrix Insights\u003c/h2\u003e\u003cp\u003eThe confusion matrices generated for each model provided additional insights into the classification performance. The ConvNeXt model and Swin model both showed minimal misclassification across the stages of ALL. The errors that did occur were mainly between adjacent stages, which is understandable given the visual similarity between these stages.\u003c/p\u003e\u003cp\u003eFor ResNet101 and VGG16, misclassifications were more frequent, particularly between the \"Early\", \"Benign\", and \"Pre\" stages. This suggests that while these models are powerful, they may struggle with finer-grained distinctions between certain stages of ALL, particularly when features are subtle.\u003c/p\u003e\u003cp\u003eThe implementation of data augmentation (random flips and rotations) and normalization was instrumental in improving model generalization and preventing overfitting. Augmentation allowed the models to train on a more diverse dataset, leading to better accuracy across models. The normalization of pixel values ensured smoother convergence during training, especially for deeper models like ResNet101. The complete comparison of the trained models on only the ALL dataset is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e below:\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eshows our first ablation study on the Acute Lymphoblastic Leukemia (ALL) dataset. The table gives a comparative overview of the accuracy of the performance and precision, recall, and F1 score of each of the tested architectures, supporting the choice of the ConvNeXt model\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eF1 Score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eConvNeXt model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e99.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSwin model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e99.39\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eResNet101\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e98.47\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVGG16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e96.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAll model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e90.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003e4.2. Discussion on Model Architectures\u003c/h2\u003e\u003cp\u003eThe ConvNeXt model excelled in this study due to its modern architecture, which effectively combines convolutional principles with the strengths of vision transformers, allowing it to recognize subtle variations in stages of Acute Lymphoblastic Leukemia (ALL). The Swin model, with its design, achieved near state-of-the-art performance, demonstrating that efficiency can coexist with complexity, making it suitable for resource-constrained clinical settings. Meanwhile, both ResNet101 and VGG16, despite being effective transfer learning models, indicated that further fine-tuning is necessary when adapting models pretrained on general datasets like ImageNet for specific medical imaging tasks; ResNet101 offered refined feature extraction but still lagged behind ConvNeXt and Swin. Lastly, the ALL model, which aimed to integrate features from various architectures, underperformed due to the under-complication of the learning process and ineffective feature integration.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003ePerformance Of The ConvNeXt Model on Extended Dataset (ALL + Lymphoma)\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eBased on the previous evaluations, ConvNeXt was chosen as the final model to be trained on the extended dataset that includes both ALL and Lymphoma cases. The reason for choosing ConvNeXt, a hybrid CNN-Transformer architecture, is that it is capable of learning hierarchical spatial features and also leverages the attention mechanism for enhanced feature extraction. The training configuration used while training the ConvNeXt model on the extended dataset is given below\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e• \u003cb\u003eInput Size\u003c/b\u003e: 224 × 224\u003c/p\u003e\u003cp\u003e• Batch Size: 32\u003c/p\u003e\u003cp\u003e• \u003cb\u003eOptimizer\u003c/b\u003e: Adam (lr = 0.001)\u003c/p\u003e\u003cp\u003e• \u003cb\u003eLoss Function\u003c/b\u003e: Cross-Entropy Loss\u003c/p\u003e\u003cp\u003e• Training Epochs: 10\u003c/p\u003e\u003cp\u003eDuring training, the ConvNeXt model converged rapidly due to its pre-trained weights and new architectural enhancements, which improved the efficiency of convolutional operations. To enhance generalization, random flipping and rotation were also used as data augmentation techniques. The training and validation accuracy and loss curve can be seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e (a) below:\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe validation accuracy was consistently above the training accuracy, which means that the model did not overfit. The new ConvNext model, which was finally trained, was quite impressive, with an accuracy of 99.72% on the test set, proving its efficacy in detecting both ALL and Lymphoma and classifying them correctly. To gauge the model's classification performance, a confusion matrix was created in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e (b) to show how the model performed in correctly classifying all the disease categories where we see that only 1 Pro ALL class was misclassified as CLL.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eExternal Dataset Evaluation of the Best Performing ConvNeXt Model\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eThe results from testing the ConvNeXt (ALL + Lymphoma) model on a completely unseen external dataset reflect the generalization ability and robustness of the model. The model reached an excellent accuracy of 97.73% on the external Lymphoma dataset. The model also got 100% accuracy on the external ALL dataset which shows that it has learned the essential features of ALL subtypes and is capable of classifying new cases. The high accuracy in both datasets shows that the model does not overfit the training data but rather generalizes the features. The outstanding performance of the model reveals its clinical significance, and thus it can be used as a reliable AI-based diagnostic tool for distinguishing between ALL and Lymphoma in real clinical practice. The confusion matrix, which shows the detailed performance of the ConvNeXt (ALL + Lymphoma) model on the ALL-external dataset, can be seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e (b), while the confusion matrix for the test of ConvNeXt (ALL + Lymphoma) on the external Lymphoma data is seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e (a)\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable border=\"1\"\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ebelow summarizes the accuracy performance of the ConvNext (ALL + Lymphoma) model on both the extended and hybrid datasets and the two external datasets.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy (%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eALL + Lymphoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e99.72\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eExternal ALL Dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e100.00\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eExternal Lymphoma Dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e97.73\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of the performance of different state-of-the-art (SOTA) models from varying research work with results from this research work, which details how our model and approach outperformed previous studies/research.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAuthors\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eAccuracy (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eDescription\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eThis study\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eALL + Lymphoma data\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eConvNeXt\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e99.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[28]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003enormal vs. malignant cells (C-NMC) dataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e91.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eUtilized CNN for ALL classification\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[29]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eWBC datasets\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eResnet-50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e90.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eClassification of Activated and Inactivated Neutrophils for Rapid Immune State Monitoring\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[30]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eALL data\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eVGG16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e98.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eUtilized VGG16 for ALL classification\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[31]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBlood samples\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSwin Transformer\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e96.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eHematology analysis framework using optical diffraction tomography\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[32]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLymphoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContent-based image retrieval system\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e98.74\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eIncorporation of deep learning with a traditional learning approach\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[33]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eALL\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eALNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e94.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eA sequential CNN-based system to predict the initial diagnosis of acute leukaemia\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[34]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eALL\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e99.31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eEfficacy of various machine learning models in the prediction of Acute Lymphoblastic Leukemia (ALL)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[35]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLymphoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eBNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e99.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eBNN for follicular lymphoma detection\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[36]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eWhole slide images - Lymphoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDLCNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e88.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eTo investigate a deep learning convolutional neural network (DLCNN) for computer-aided\u003c/p\u003e\u003cp\u003edetection of MCL\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[37]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLymphoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e95.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eUse Deep Learning with a convolutional neural network (CNN) algorithm to build a lymphoma diagnostic model\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAll Model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e90.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo illustrate the performance of our suggested framework, we compared our results with some state-of-the-art (SOTA) approaches from the recent literature (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The performance of our unified ConvNeXt model, based on the compound dataset (99.72% accuracy), is also better or at least competitive with that of specialized models specific to ALL or lymphoma. For example, our model outperforms the CNN model by [28] (91.10%) and the ALNet system [33] (94.50%) in detecting ALL. Our single-model-based two-disease detection scheme has been proven effective and innovative on independent datasets.\u003c/p\u003e\u003c/div\u003e"},{"header":"Conclusion and Recommendations","content":"\u003cp\u003eThis research has effectively developed a sophisticated AI baseline diagnostic framework for detecting and predicting Acute Lymphoblastic Leukemia (ALL) and Lymphoma using a single model within deep learning frameworks. A comparative analysis of multiple models was initially performed, and the ConvNeXt model was selected as the most efficient among Swin Transformers, ResNet101, VGG16, and a Custom CNN, achieving an accuracy of 99.69% on the ALL dataset. The ConvNext model was then used to train a multiple dataset comprising both the ALL and Lymphoma datasets, where it achieved 99.72% on the extended ALL + Lymphoma dataset. The model was tested on external, unseen datasets, where it achieved 100% accuracy on the ALL dataset and 97.73% accuracy on the Lymphoma dataset, demonstrating the model's generality and potential for real-life use. These results support the idea that combination architectures like ConvNeXt that incorporate both convolutional and transformer-based features offer better feature learning and classification performance in the field of hematological image comprehension.\u003c/p\u003e\u003cp\u003eThe study also addressed important issues in the application of AI in medical diagnostics, such as small dataset issues, overfitting, and computational complexity. In order to guarantee the robustness and scalability of the model for clinical practice, the research incorporated data augmentation, external validation, and balanced dataset techniques.\u003c/p\u003e\u003cp\u003eThis is a significant breakthrough in the detection of hematological cancer, and this AI driven diagnostic system has a very high performance. It can potentially assist pathologists and oncologists in the early diagnosis and classification of leukemia and lymphoma with a single model and therefore lead to better patient outcomes and treatment rates.\u003c/p\u003e\u003cp\u003eFuture work should further extend the dataset to detect more diseases thus making it more robust, as well as the integration of this system into the healthcare as a real time diagnostic tool in clinical practice. This study introduces the application of AI in multiple blood cancer detection in hematological oncology and therefore represents a starting point for the development of fully automated, highly accurate, and easy to use cancer diagnostic systems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eEthical Statement\u003c/h2\u003e\u003cp\u003eThis study does not contain any studies with human or animal subjects performed by any of the authors.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003ch2\u003eConflicts of Interest\u003c/h2\u003e\u003cp\u003eThe authors declare that they have no conflicts of interest to this work.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAcknowledgments:\u003c/h2\u003e\u003cp\u003eThe authors would like to thank the University of Salford for providing the funding for the article publication costs.\u003c/p\u003e\u003ch2\u003eData Availability Statement\u003c/h2\u003e\u003cp\u003eThe data that support the findings of this study as well as the analysis and model training are openly available in the repository \u0026ldquo;ALL-Detection\u0026rdquo; specifically in the notebook file named \u0026ldquo;lymphoma-leukemia-detection.ipynb\u0026rdquo; accessed via the link lymphoma-leukemia-detection\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eKwee TC, Kwee RM, Nievelstein RAJ (2008) Imaging in staging of malignant lymphoma: A systematic review. Blood 111:1072\u0026ndash;1082\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMalard F, Mohty M (2020) Acute lymphoblastic leukaemia. Lancet 395:1146\u0026ndash;1162\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCoustan-Smith E, Mullighan CG, Onciu M et al (2011) New markers for minimal residual disease detection in acute lymphoblastic leukemia. Blood 117:6267\u0026ndash;6276\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShalabi H, Angiolillo AL, Fry TJ et al (2020) Disease detection methodologies in relapsed B-cell acute lymphoblastic leukemia: Opportunities for improvement. Pediatr Blood Cancer 67:e28149\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWijetunga NA, Imber BS, Caravelli JF, Mikhaeel NG (2021) Radiomics in lymphoma: A systematic review. Br J Radiol 94:20210285\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePrasad P, Anbarasi LJ (2024) Acute Lymphoblastic Leukemia Subtypes Detection using Vision Transformer Model. In: Proceedings of the 5th International Conference on Data Intelligence and Cognitive Informatics, ICDICI 2024. Institute of Electrical and Electronics Engineers Inc., 1413\u0026ndash;1418\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlbeeshi AA, Alshanbari HS (2023) Modeling of the Acute Lymphoblastic Leukemia Detection by Convolutional Neural Networks (CNNs). Curr Med Imaging Rev 19:835\u0026ndash;844\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIbtekar T, Haque MR, Srizon AY (2023) Comparative Study Between ResNet And EfficientNet Family For Classification of Leukemia. In: 2023 26th International Conference on Computer and Information Technology (ICCIT). IEEE, 1\u0026ndash;6\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRezayi S, Mohammadzadeh N, Bouraghi H, Saeedi S, Mohammadpour A (2021) Timely Diagnosis of Acute Lymphoblastic Leukemia Using Artificial Intelligence-Oriented Deep Learning Methods. Comput Intell Neurosci 2021:5478157\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbunadi I, Senan EM (2022) Multi-Method Diagnosis of Blood Microscopic Sample for Early Detection of Acute Lymphoblastic Leukemia Based on Deep Learning and Hybrid Techniques. Sens (Basel) 22:1629\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMasilamani V, Devanesan S, AlSalhi MS, AlQahtany FS, Farhat KH (2020) Fluorescence spectral detection of acute lymphoblastic leukemia (ALL) and acute myeloid leukemia (AML): A novel photodiagnosis strategy. Photodiagnosis Photodyn Ther 29:101634\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhandekar R, Shastry P, Jaishankar S, Faust O, Sampathila N (2021) Automated blast cell detection for Acute Lymphoblastic Leukemia diagnosis. Biomed Signal Process Control 68:102690\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGenovese A, Hosseini MS, Piuri V, Plataniotis KN, Scotti F (2021) Histopathological transfer learning for acute lymphoblastic leukemia detection. In: CIVEMSA 2021 - IEEE International Conference on Computational Intelligence and Virtual Environments for Measurement Systems and Applications, Proceedings. Institute of Electrical and Electronics Engineers Inc\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNautiyal U, Bhatt A, Chauhan R, Rawat R, Gupta R (2024) A Modified Conventional Neural Network for Detecting and Classifying Leukemia. In: Proceedings of the 2nd International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics, ICIITCEE 2024. Institute of Electrical and Electronics Engineers Inc\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlbano D, Bruno A, Patti C, Micci G, Midiri M, Tarella C, Galia M (2019) Whole-body magnetic resonance imaging (WB-MRI) in lymphoma: State of the art. Hematol Oncol 38:12\u0026ndash;21\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWeisman AJ, Kieler MW, Perlman SB, Hutchings M, Jeraj R, Kostakoglu L, Bradshaw TJ (2020) Convolutional neural networks for automated pet/ct detection of diseased lymph node burden in patients with lymphoma. Radiol Artif Intell 2:e200016\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSteinbuss G, Kloss M, Laleh G, Wagner M, Zgorzalski W, Bol K, Lapi E, Lengerke C, Tatsch T, M\u0026uuml;ller A (2021) Deep learning for the classification of non-hodgkin lymphoma on histopathological images. Cancers (Basel) 13:2419\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSheng B, Zhou M, Hu M, Li Q, Sun L, Wen Y (2020) A blood cell dataset for lymphoma classification using faster R-CNN. Biotechnol Biotechnol Equip 34:413\u0026ndash;420\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKanavos T, Birbas E, Zanos TP (2025) A Systematic Review of the Applications of Deep Learning for the Interpretation of Positron Emission Tomography Images of Patients with Lymphoma. Cancers (Basel) 17:69\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMiyoshi H, Maeshima AM, Nomoto J, Tng N, Tsuruoka S, Kono S, Nakai T, Fukuhara S, Munakata W, Maruyama D, Tobinai K, Kobayashi Y, Kober T (2020) Deep learning shows the capability of high-level computer-aided diagnosis in malignant lymphoma. Lab Invest 100:1300\u0026ndash;1310\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang YF, Zhao E, Shi Y, Zhang H, Yang YY (2024) Multicenter investigation of preoperative distinction between primary central nervous system lymphomas and glioblastomas through interpretable artificial intelligence models. Neuroradiology 66:1765\u0026ndash;1774\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBai A, Si M, Xue P, Qu Y, Jiang Y (2024) Artificial intelligence performance in detecting lymphoma from medical imaging: a systematic review and meta-analysis. BMC Med Inf Decis Mak 24:14\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGhaderzadeh M, Aria M, Hosseini A, Asadi F, Bashash D, Abolghasemi H (2022) A fast and efficient CNN model for B-ALL diagnosis and its subtypes classification using peripheral blood smear images. Int J Intell Syst 37:5113\u0026ndash;5133\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOrlov NV, Chen WW, Eckley DM, Macura TJ, Shamir L, Jaffe ES, Goldberg IG (2010) Automatic classification of lymphoma images with transform-based global features. IEEE Trans Inf Technol Biomed 14:1003\u0026ndash;1013\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaggle Multi Cancer Dataset. Available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.kaggle.com/datasets/obulisainaren/multi-cancer-dataset\u003c/span\u003e\u003cspan address=\"https://www.kaggle.com/datasets/obulisainaren/multi-cancer-dataset\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed 26 July 2024\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J: Deep Residual Learning for Image Recognition. In: IEEE Conference on Computer Vision and, Recognition P (2016) (CVPR). IEEE, 770\u0026ndash;778, 2016\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang MH, Chong KKL, Lin Z, Yu X, Pan Y (2023) An Explainable Artificial Intelligence-Based Robustness Optimization Approach for Age-Related Macular Degeneration Detection Based on Medical IOT Systems. Electron (Switzerland) 12:2697\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eUllah MZ, Zheng Y, Song J, Aslam S, Xu C, Kiazolu GD, Wang L (2021) An Attention-Based Convolutional Neural Network for Acute Lymphoblastic Leukemia Classification. Appl Sci (Basel) 11:10662\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang X, Jeon H, Liu J, Yao J, Wei M, Han W, Chen J, Sun L, Han J (2021) Deep-Learning Based Label-Free Classification of Activated and Inactivated Neutrophils for Rapid Immune State Monitoring. Sens (Basel) 21:512\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlim MS, Bappon SD, Sabuj SM, Islam MJ, Tarek M, Azam MS, Islam MM (2024) Integrating convolutional neural networks for microscopic image analysis in acute lymphoblastic leukemia classification: A deep learning approach for enhanced diagnostic precision. Syst Soft Comput 6:200121\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRyu D, Bak TH, Ahn D, Kang HY, Oh S, Min HS, Lee S, Lee J (2023) Deep Learning-Based Label-Free Hematology Analysis Framework Using Optical Diffraction Tomography. Heliyon 9:e18297\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eReena MR, Ameer PM (2022) A content-based image retrieval system for the diagnosis of lymphoma using blood micrographs: An incorporation of deep learning with a traditional learning approach. Comput Biol Med 145:105463\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBold\u0026uacute; L, Merino A, Acevedo A, Molina A, Rodellar J (2021) A deep learning model (ALNet) for the diagnosis of acute leukaemia lineage using peripheral blood cell images. Comput Methods Programs Biomed 202:105999\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eV KD SG (2024) M G, M KS, A MC, S R: Advanced Machine Learning Techniques for Accurate Detection and Diagnosis of Acute Lymphoblastic Leukemia (ALL). In: 2024 5th International Conference on Electronics and Sustainable Communication Systems (ICESC). IEEE, 1362\u0026ndash;1366\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSyrykh C, Abreu A, Amara N, Siegfried A, Maisongrosse V, Frenois FX, Martin L, Rossi C, Laurent C, Brousset P (2020) Accurate diagnosis of lymphoma on whole-slide histopathology images using deep learning. NPJ Digit Med 3:63\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou Z, Jain P, Lu Y, Macapinlac H, Wang ML, Son JB, Pagel MD, Xu G, Ma J (2021) Computer-aided detection of mantle cell lymphoma on 18F-FDG PET/CT using a deep learning convolutional neural network. Am J Nucl Med Mol Imaging 11:260\u0026ndash;270\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAchi HE, Belousova T, Chen L, Wahed A, Wang I, Hu Z, Kanaan Z, Rios A, Nguyen AND (2019) Automated Diagnosis of Lymphoma with Digital Pathology Images Using Deep Learning. Ann Clin Lab Sci 49:153\u0026ndash;160\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Salford","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Acute Lymphoblastic Leukemia (ALL), Deep Learning, Peripheral Blood Smear (PBS), Lymphoma, Convolutional Neural Networks (CNNs)","lastPublishedDoi":"10.21203/rs.3.rs-8155029/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8155029/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAcute lymphoblastic leukemia (ALL) and Lymphoma are important diseases that need to be detected early, but detection with traditional methods may be slow and inconsistent. Although current AI methods present potential advantages, they are frequently limited by their reliance on small input data, overfitting, and a lack of external verification, which are unfavorable for their clinical implementation. This paper proposes a unified framework for creating a shared deep learning model for dual disease detection, based on the ConvNeXt design. ConvNeXt's superiority was demonstrated in a first ablation, wherein the Model was compared to Swin Transformer, ResNet101, VGG16, and a custom CNN on an ALL dataset, achieving a best accuracy of 99.69%. The selected ConvNeXt model was then optimized and retrained on a larger dataset consisting of both ALL and Lymphoma samples. As a result, there was only one highly consistent diagnostic tool that attained an accuracy of 99.72% on the mixed test set. More importantly, the practical utility of the framework in obtaining significant results was validated through extensive testing using completely new, previously unseen external datasets. The framework demonstrated an outstanding degree of generalizability, achieving 100% accuracy on an independent ALL dataset and 97.73% on an independent Lymphoma dataset. This work presents a well-represented, fully automated, and externally validated hematological diagnosis system that demonstrates a possible route of implementing trusted AI in direct care practice.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e","manuscriptTitle":"Deep Learning Models for Accurate Leukemia \u0026amp; Lymphoma Detection","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-24 05:19:17","doi":"10.21203/rs.3.rs-8155029/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"54459608-aa5a-43b8-9150-50b960cfef36","owner":[],"postedDate":"November 24th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":58251723,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-11-24T05:19:18+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-24 05:19:17","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8155029","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8155029","identity":"rs-8155029","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}