Advancing Breast Cancer-AI Diagnostics: An Explainable Deep Learning Model Using 2D Grayscale Ultrasound Imaging

Advancing Breast Cancer-AI Diagnostics: An Explainable Deep Learning Model Using 2D Grayscale Ultrasound Imaging | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (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];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Advancing Breast Cancer-AI Diagnostics: An Explainable Deep Learning Model Using 2D Grayscale Ultrasound Imaging View ORCID Profile Ghulam Husain Abbas , Shafee Ur Rehman , Ghazal Bargshady doi: https://doi.org/10.1101/2025.10.31.25339239 Ghulam Husain Abbas 1 Faculty of Medicine, Ala-Too International University , 1/8 Ankara St, Bishkek, Kyrgyzstan MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Ghulam Husain Abbas For correspondence: ghulamhusainabbas.abbas{at}alatoo.edu.kg Shafee Ur Rehman 1 Faculty of Medicine, Ala-Too International University , 1/8 Ankara St, Bishkek, Kyrgyzstan PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ghazal Bargshady 2 Faculty of Science & Technology, University of Canberra , ACT, 2617, Australia PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Background Breast cancer stands as the primary reason for fatality in female patients from cancer worldwide. The diagnostic precision of ultrasound imaging depends on operator skills because it lacks invasive procedures but remains accessible through wide availability. The medical field needs automated systems which provide explainable results to help doctors achieve better breast lesion diagnosis accuracy. Objectives This research presents a new deep learning diagnostic system based on MobileNetV2 architecture which achieves optimal results for analyzing 2D grayscale (B-mode) ultrasound images. The system combines Grad-CAM explainability tools with clinical-grade functionality to generate interpretable results for both correct and incorrect predictions. Methods The BUSI dataset served as the basis for model training and evaluation because it contained a total of 780 ultrasound images that contains 437 benign images, 210 malignant images and 133 normal images. We divided the data into training and validation sets which made up 80% and 20% of the total data respectively. We then standardized and normalized all images by resizing them to 192×192 pixels and converting them into 3-channel images. The model design included frozen base layers, global average pooling, dropout regularization and a fully connected classification head. The model received 7 training runs with a total of 150 epochs with 16 samples per batch using adam optimizer and categorical cross-entropy as the loss function. Results The proposed model achieved 90.1% validation accuracy, 0.90 precision, 0.90 recall, 0.90 F1-score and 0.88 macro-averaged AUC. The Grad-CAM visualizations showed precise localization of features within the lesion area. The t-SNE projection of learned features displayed three separate lusters which corresponded to malignant, benign and normal tissue types. The system produced most of its errors when it failed to distinguish between benign and malignant lesions that appeared very similar to each other. Conclusion The system provides both high diagnostic precision and clinical understanding which represents a major breakthrough for ultrasound-based breast cancer screening systems. Further research will direct upcoming work toward validating the system externally and implementing real-time functionality, obtaining necessary regulatory approvals as well as inclusion for multi-modality in deep learning. 1. Introduction Breast cancer stands as the leading cancer type which affects women worldwide because it causes 2.3 million new cases and 685,000 deaths annually according to the World Health Organization (WHO) [ 1 ]. The implementation of early detection methods for breast cancer reduction depends on traditional diagnostic methods which encounter multiple challenges regarding accessibility and precision [ 2 , 3 , 4 ]. The current population screening standard mammography shows decreased performance in dense breast tissue patients while its availability remains restricted in underserved areas [ 5 ]. MRI provides excellent sensitivity but its high cost and screening impracticality make it unsuitable for regular breast cancer screening [ 6 ]. The medical field now uses ultrasound imaging as a practical non-invasive radiation-free diagnostic tool because it provides easy access to patients [ 7 ]. The imaging method proves beneficial for younger patients and helps doctors evaluate mammography-detected masses [ 8 , 9 ]. The clinical value of breast ultrasound encounters three main obstacles which stem from operator-dependent results and subjective readings and inconsistent lesion appearances [ 10 , 11 ]. Deep learning technology within Artificial Intelligence demonstrates exceptional potential to automate medical image interpretation for different healthcare applications [ 12 , 13 , 14 ]. For other DL methods John et al. (2022), applied Alexnet in CT-Scan image he recived 98%, which shows that it is possible to reach almost 100% accuracy. The current AI solutions for breast ultrasound diagnosis encounter three major obstacles which include unexplained decision-making processes that lower medical staff confidence and regulatory approval rates and insufficient model detail that treats benign and malignant lesions as equal categories while ignoring normal breast tissue as a distinct diagnostic group and insufficient analysis of incorrect classifications for safe clinical implementation [ 15 , 16 , 17 ]. The research develops an explainable AI system which utilizes MobileNetV2 convolutional neural networks to solve current AI system deficiencies. The system uses 2D grayscale ultrasound images and Grad-CAM for visual interpretability through its design. The system delivers precise diagnostic explanations for correct and incorrect dentification results in addition to achieving high classification accuracy. The system represents a major advancement toward clinical AI transparency because it moves beyond uninterpretable "black-box" models. The research presents a novel approach by developing a three-class classification system for 2D ultrasound images with visual explanations through a lightweight yet effective architecture. The study delivers specific explanations for each case of incorrect classification. The research demonstrates model discriminability through t-SNE feature space clustering analysis. The research presents performance evaluation results of multiple peer-reviewed deep learning models through tabular comparison. The research follows an original article structure which enables reproducibility through methodological clarity and technical precision and meets clinical translation requirements. 1.1 Contributions of this Research This research makes three major contributions to the field of AI-driven breast cancer diagnostics. First, it introduces a novel lightweight deep learning framework based on MobileNetV2 that performs three-class classification of breast lesions, benign, malignant, and normal, using 2D grayscale ultrasound images with high accuracy and computational efficiency suitable for real-time applications. Second, it integrates Gradient-weighted Class Activation Mapping (Grad-CAM) to provide transparent and clinically interpretable visual explanations, enabling medical professionals to understand the reasoning behind both correct and incorrect model predictions. Third, it presents comprehensive performance validation and benchmarking against established deep learning architectures, demonstrating superior diagnostic accuracy, strong feature separability through t-SNE visualization, and robust generalization potential. Together, these contributions advance the development of explainable, efficient, and deployable AI systems for breast cancer screening in resource-limited clinical environments. 2. Methods and Materials 2.1 Introduction This section outlines the methodological framework and experimental design used to develop and evaluate the proposed explainable deep learning system for breast cancer classification. It details each stage of the research process, including dataset selection and preprocessing, model architecture, training strategy, evaluation metrics, and explainability analysis. The study utilizes a publicly available breast ultrasound dataset to ensure reproducibility and ethical transparency, while employing a lightweight MobileNetV2 architecture optimized for accuracy and computational efficiency. Preprocessing and augmentation techniques were applied to enhance model generalization, and multiple performance metrics were used to comprehensively assess classification reliability. Additionally, interpretability was achieved through the integration of Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize decision-relevant regions within ultrasound images. Together, these methodological components form a robust foundation for developing an accurate, interpretable, and clinically applicable AI model for breast cancer diagnosis. 2.2 Dataset The research used the Breast Ultrasound Images (BUSI) dataset from Al-Dhabyani et al. [ 18 ] which contains 780 2D grayscale ultrasound images that fall into three diagnostic groups: normal (133 images), benign (437 images) and malignant (210 images). The images are labeled by expert radiologists and stored in PNG format with different resolutions. The images were distributed into two distinct folders for training and validation/testing purposes. We applied stratified sampling to balance the dataset so that each class maintained equal representation during training. The labels received one-hot encoding for multi-class classification purposes. 2.3 Ethical Statement The research data consists of public domain images that have been completely de-identified and contain no personal or sensitive information. The research study works with no human participants so it does not create any ethical or privacy concerns. The research meets standard criteria for secondary data analysis because it does not require Institutional Review Board (IRB) approval or ethical review. 2.4 Data Preprocessing The input images underwent preprocessing to achieve compatibility with MobileNetV2 architecture requirements. The images received a transformation to 192 × 192 pixels and acquired three-channel format for network input compatibility. The preprocessing process involved two steps: normalization, which transformed pixel values to the range (0 & 1), and data augmentation methods applied to training images for better model performance. The training images received multiple data augmentation operations, which included random horizontal and vertical flipping and rotation between 0° and 20° and zoom adjustments up to 10% and brightness and contrast modifications. The dataset received stratified sampling for partitioning into 80% training data and 20% testing validation data to preserve class equilibrium. 2.5 Model Architecture The research employed transfer learning with MobileNetV2 as its base feature extractor because this lightweight CNN model provides both high accuracy and efficient computation. The pre-trained MobileNetV2 model received its ImageNet training before we removed its classification layers for subsequent use. The architecture received modifications through the following steps: The input dimensions were set to (192, 192, 3) and the base model layers were frozen to maintain learned representations before adding a new classification head that included GlobalAveragePooling2D followed by Dense with 128 units and ReLU activation and Dropout with rate 0.3 for regularization and a final Dense layer with three units and softmax activation for multi-class classification. The designed architecture delivers high accuracy performance while keeping computational requirements low which makes it suitable for mobile and point-of-care ultrasound diagnostic systems, a simple diagram of the proposed framework is in Figure 11 . 2.6 Training Details The model received compilation with Adam optimizer at 0.0001 learning rate to achieve stable and efficient training results. The model used categorical cross-entropy as its loss function because it suits multi-class classification problems. The training process ran for seven runs with a total of 150 epochs with a batch size of 16 which resulted from early stopping analysis and observation of convergence patterns during experimental testing. The training process received stability through two callback functions which included EarlyStopping with patience=3 and restore_best_weights=True for stopping training then validation performance stopped improving and ReduceLROnPlateau with factor=0.5 and patience=2 for automatic learning rate adjustments based on validation loss stagnation, Table 7 shows the key hyperparameters of the model. 2.7 Evaluation Metrics The proposed model was evaluated using a hold-out validation strategy rather than cross-validation. The dataset was divided into training and validation subsets in an 80:20 ratio using the train_test_split() method, ensuring that the test data remained completely unseen during training. The model’s performance was assessed through a comprehensive set of quantitative and visual metrics, including accuracy, precision, recall, F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). These metrics were calculated for each class as well as macro-averages to evaluate the classification stability and balance across all categories. The confusion matrix presented in Figure 5 illustrates both correct and incorrect predictions for each class, while Figure 7 displays class-wise prediction accuracy and misclassification patterns. Additionally, the detailed per-class misclassification analysis is provided in Figure 10 , offering deeper insight into the model’s discriminative ability and generalization performance. 2.8 Explainability The model interpretability received improvement through the application of Gradient-weighted Class Activation Mapping (Grad-CAM), which identifies specific areas in ultrasound images that distinguish between classes. We used Grad-CAM to analyze all test images after the model finished its predictions so we could see which image parts led to the model’s decisions. The heatmaps received classification based on the prediction results, which included true positives (TP) and false positives (FP) and false negatives (FN), and true negatives (TN) for error analysis purposes, these results of the Grad-CAM are available to view in Figure 4 . We evaluated clinical relevance through an analysis of whether the activated areas matched important diagnostic indicators, which include hypoechoic masses and irregular borders, and posterior acoustic shadowing. The Grad-CAM visualizations in Figure 4 demonstrate different categories of the model’s output and examples of real-time results are available as supplementary data (normal.pdf, benign.pdf, and malignant.pdf). The model’s learned feature space became visible through a t-distributed stochastic neighbor embedding (t-SNE) plot ( Figure 2 ), which showed distinct patterns of separation between different diagnostic classes. 3. Results 3.1 Dataset Composition and Distribution We used 780 breast ultrasound images to create the final dataset, which we divided into three diagnostic categories: Benign (N = 437), Malignant (N = 210), and Normal (N = 133). The class distribution in Figure 1 shows that benign cases make up the largest group, while malignant and normal cases follow in order of frequency. The preprocessed images received uniform resolution of 192 × 192 × 3 to match deep learning architecture requirements of MobileNetV2 while maintaining the original 2D grayscale format. Download figure Open in new tab Figure 1. Image Count per Class (Train vs Validation) shows distribution of images across benign, malignant, and normal classes for both training and validation datasets. Download figure Open in new tab Figure 2. t-SNE Visualization of Model Feature Space shows two-dimensional t-SNE plot showing feature clustering learned by the model for each class, indicating separability of benign, malignant, and normal samples. Download figure Open in new tab Figure 3. Training and Validation Performance over Epochs (A) Accuracy and (B) loss trends across epochs, showing model convergence and consistent improvement during training and validation. Download figure Open in new tab Figure 4. Example Activation Maps for True and Misclassified Samples visualization of true positive (TP), false positive (FP), and false negative (FN) samples to highlight regions influencing model predictions. Download figure Open in new tab Figure 5. Confusion Matrix illustrating the number of correct and incorrect predictions for each class, with stronger diagonal values indicating higher classification accuracy. 3.2 Training Performance The MobileNetV2 model trained on BUSI data reached 96% training accuracy and 90.1% validation accuracy during 150training epochs. The training and validation accuracy and loss curves in Figure 3a and 3b show steady convergence and early stabilization, which indicates proper learning behavior. The model avoided overfitting because it used dropout and data augmentation and early stopping techniques for optimal performance on new data. 3.3 Evaluation Metrics The model achieved 90.1% accuracy on the test set containing 161 images according to Table 1 and Figure 3a . The model showed excellent ability to recognize new data because it achieved 90.1% accuracy in the test set compared to prior models where they fall between 64-82% accuracy. The model demonstrates strong diagnostic performance through its ability to distinguish between benign and malignant and normal cases as shown in Table 1 with complete precision and recall and F1-score values. The model demonstrates high performance in identifying benign and malignant and normal breast ultrasound images according to the results. The macro-averaged F1-score was 0.86 , and the weighted average reached 0.90 , affirming strong class balance despite slight performance drop in the "Normal" category. Detailed precision, recall, and F1-score for each class are presented in Table 1 . Figure 9 illustrates the prediction confidence distribution across all test samples. The model demonstrates high confidence levels for correctly classified cases and moderate confidence in borderline examples, indicating effective calibration and reliable probabilistic output. View this table: View inline View popup Download powerpoint Table 1. Performance Metrics per Class on tested dataset (20% tested and 80% trained) 3.4 Confusion Matrix and Misclassification The test set prediction results from the model appear in Figure 5 through a confusion matrix that shows class-by-class performance. A summarized confusion matrix of true and predicted classifications is shown in Table 2 , highlighting the model’s class-wise distribution of correct and incorrect predictions. The model experienced most of its prediction errors when distinguishing between Normal and Benign cases because these two categories share similar tissue patterns and minimal visual distinctions in breast ultrasound images. The model’s prediction process and confusion sources become clearer through Figure 8 which displays representative misclassified samples with Grad-CAM overlays that show the areas the model used for prediction. A comprehensive breakdown of misclassified cases and the clinical interpretations is provided in Table 6 , including false positives and false negatives along with the diagnostic impact and suggested mitigations. View this table: View inline View popup Download powerpoint Table 2. Confusion Matrix Summary View this table: View inline View popup Download powerpoint Table 3. Comparison with Previous Breast Ultrasound AI Models 3.5 Receiver Operating Characteristic (ROC) Analysis The multiclass ROC curves for the test set appear in Figure 6 . The model achieved excellent discriminative performance through its Area Under the Curve (AUC) scores which reached 0.88 for Benign and 0.86 for Malignant and 0.97 for Normal classes. The model shows excellent class separation through its high AUC scores which prove its ability to perform well across all three diagnostic groups. 3.6 Explainability via Grad-CAM The application of Grad-CAM to test cases in supplementary data (normal.pdf, benign.pdf and malignant.pdf) enhances model interpretability. The model bases its predictions on lesion boundaries and echotexture patterns through the highlighted discriminative regions shown in true positive and false positive samples. The model uses image features that experts in radiology use for diagnosis because the highlighted areas match their assessment criteria. 3.7 Individual Patient-Level Outputs The deployed model produced three diagnostic reports from 3 different individual patient open-sourced data that is available online [ 18 ], the diagnostic reports are available as supplementary files. The reports present the model’s diagnosis output along with the original ultrasound image and Grad-CAM heatmap and class probability values, and a human-readable explanation of the prediction process. The detailed reports demonstrate how the model works in clinical environments because they provide complete information that helps non-specialist healthcare providers understand AI-generated results. 3.8 Additional Feature and Impact Analysis The analysis in Figures 9 and 10 shows the extent of misclassification errors together with Grad-CAM feature weight analysis, which reveals the specific image areas that led to incorrect predictions. The t-SNE visualization in Figure 2 shows how the model separates diagnostic classes through its 2D feature representation of high-dimensional learned data. The analysis provides both numerical and visual evidence that the network detects important features and recognizes areas where it makes mistakes. Download figure Open in new tab Figure 6. Multiclass ROC curves for each class with corresponding Area Under Curve (AUC) values showing model discriminative performance. Download figure Open in new tab Figure 7. Class-wise Precision, Recall, and F1-scores showing comparison of precision, recall, and F1 metrics across benign, malignant, and normal classes, summarizing per-class classification performance. Download figure Open in new tab Figure 8. Misclassified Samples per Class shown with a bar chart displaying the total number of misclassified samples for each class, indicating relative error frequency. Download figure Open in new tab Figure 9. Prediction Confidence Distribution shows a histogram of predicted probabilities showing the confidence levels of the model for each class. Download figure Open in new tab Figure 10. Misclassifications per Class heatmap showing specific misclassification patterns between true and predicted classes, highlighting confusion areas in the model. Download figure Open in new tab Figure 11. A general diagram of the proposed framework for visual ease 3.9 Model Comparison and Benchmarking The proposed model outperforms ResNet50 and EfficientNetB0 in all three evaluation criteria of accuracy and computational efficiency and interpretability according to Tables 4 . Table 5 summarizes the clinical readiness and real-world deployment potential of the proposed system, demonstrating its lightweight design, rapid inference time, and explainability features that support integration into mobile or low-resource clinical environments. The system combines an efficient MobileNetV2 backbone with Grad-CAM visualizations and PDF-based diagnostic outputs to provide black-box-free AI solutions. The model demonstrates excellent performance through its accuracy rate above 88% while providing transparent results and user-friendly interfaces which make it ready for clinical deployment. A detailed quantitative comparison between our model and prior CNN architectures is presented in Table 4 , demonstrating its superior validation accuracy and real-time deployment capability View this table: View inline View popup Download powerpoint Table 4. Evidence-Based Comparison Showing Superiority of Proposed Model View this table: View inline View popup Download powerpoint Table 5. Clinical Readiness and Real-World Impact of the Proposed AI System View this table: View inline View popup Table 6. Detailed Error Analysis of Misclassifications with Clinical Contextual Insights View this table: View inline View popup Download powerpoint Table 7. Model Key Hyperparameters 4. Discussion 4.1 Summary of Key Findings The research presents a complete explainable lightweight deep learning system which detects multiple breast cancer types from 2D grayscale ultrasound images [ 18 ]. The proposed system uses MobileNetV2 as its backbone to achieve efficient image classification through transfer learning from pre-trained models for malignant and benign and normal tissue identification [ 19 ]. The network processed standardized images through preprocessing to achieve both high-quality feature extraction and fast processing times for real-time implementation. The model achieved 90.1% accuracy when tested on the validation dataset [ 20 ]. The model achieved high precision and recall and F1-scores for each diagnostic category which indicates its balanced performance across all three classes [ 21 ]. The system demonstrated outstanding class separation through its macro-average AUC score of 0.95 which proves its ability to detect small variations in breast tissue patterns [ 22 ]. The model provides both quantitative results and clear explanations to medical professionals. The system produced Grad-CAM visualizations for each test image to show which areas influenced the prediction results. The heatmaps showed direct correspondence with essential diagnostic elements which included lesion edges and echotexture details and posterior acoustic shadows thus enabling radiologists to verify the model’s decision-making process. The t-SNE analysis used high-dimensional learned features to create a 2D representation which showed distinct clusters between malignant and benign and normal cases thus proving the network’s ability to detect discriminative patterns and reduce class overlap [ 23 ]. The model demonstrates its ability to generalize to new data through both qualitative and quantitative results from these analyses. The proposed system outperforms existing solutions through its real-time deployment capability and its ability to generate patient-specific Grad-CAM reports and PDF outputs that provide clinical interpretability. The framework achieves both diagnostic reliability and practical clinical application through its combination of high performance and transparent and user-friendly features which benefit resource-constrained and non-specialist medical environments. 4.2 Interpretation of Results The model demonstrated high recall performance for both benign and malignant cases, which indicates its ability to identify actual positive instances correctly. The model shows most classification errors between malignant and benign images because it faces the typical challenge of distinguishing small irregular, or posteriorly shadowed lesions in B-mode scans. The model demonstrates human-level diagnostic challenges because it produces errors that match real-world difficulties in medical imaging interpretation. The model produced most of its false negative results by identifying low-contrast malignant lesions as normal tissue [ 24 ]. The detection of both human and algorithmic signals becomes more difficult when grayscale sonography shows minimal echotexture differences and poor lesion visibility. The clinical deployment of this system requires awareness about its detection limitations because it needs careful evaluation of cases with uncertain or faint imaging characteristics. The t-SNE visualization shows that the network learned distinct features for each diagnostic class because its internal representations stay separate between the three categories. The model used Grad-CAM heatmaps to show its consistent focus on relevant clinical areas, which included lesion borders and echoic centers for making accurate predictions. The interpretability studies show that the model focuses on areas which match radiological standards, thus establishing its potential for clinical use 4.3 Comparison with Prior Work As seen in Table 3 , our model outperformed heavier architectures by optimizing for speed, accuracy, and explainability . Moreover, unlike many black-box CNNs, this system provides transparency in decisions , vital for clinical adoption. 4.4 Clinical Relevance The clinical application of ultrasound imaging for breast lesion evaluation benefits from its accessibility and low cost, and non-ionizing nature. The diagnostic results from ultrasound imaging depend on operator experience because it leads to different interpretation outcomes. The proposed explainable AI model functions as a decision-support system, which would benefit low-resource medical facilities because they lack sufficient experienced radiologists. The system provides visual Grad-CAM heatmaps that help clinicians build trust in the system while teaching them to improve their ability to detect lesions. The model operates efficiently because of its lightweight design, which allows it to run on mobile devices and bedside equipment for point-of-care diagnostic needs. The portable nature of this system enables advanced diagnostic support to reach remote locations with limited resources, which results in fast and dependable breast cancer assessments that maintain their clinical value and interpretability. 4.5 Strengths The system provides both high explainability and efficiency through Grad-CAM heatmaps, which show the specific image areas that influence model predictions. The model uses MobileNetV2 architecture to achieve fast processing times for real-time applications, while its training methods with stratified data splits and augmentation and balanced evaluation metrics produce reliable performance across different case types. The system provides both high performance and clear explanations of its operations. The system provides three-class classification for breast lesions, which includes malignant and benign, and normal categories instead of binary output. The system provides enhanced clinical value through its ability to classify all three breast lesion categories. The t-SNE visualizations of learned features show distinct class clusters and abstracted features, which help explain the decision-making process and build trust for practical applications. 4.6 Limitations The research depends on the BUSI dataset for its results because it uses data from a single centre. The model needs external validation to prove its ability to work with different patient groups and various ultrasound equipment and imaging procedures. The study depends only on grayscale B-mode images for analysis but does not include Doppler or elastography, or multimodal data, which could enhance diagnostic capabilities. The dataset contains insufficient annotation detail because it lacks pixel-level segmentation and pathologically confirmed labels, which restricts the training data’s clinical precision and detailed analysis. 4.7 Future Work This research needs to validate its results through multi-centre studies using real-world clinical ultrasound images to prove its effectiveness with different patient groups and imaging equipment. The system’s predictive accuracy will improve when it incorporates electronic health records and patient demographic information to generate individualized diagnostic results. The research should focus on three main objectives which include obtaining regulatory clearance for clinic employment and developing active learning methods to improve model performance through continuous case additions as well as inclusion for multi-modality in deep learning for increased accuracy and validation. A real-time diagnostic system that works with handheld ultrasound probes would provide extended accessibility for breast cancer screening at clinical sites and remote locations. Integration with clinical workflow systems such as PACS or mobile ultrasound devices will be explored in future iterations as well as inclusion of multi-modality with other systems. 5. Conclusion Our research has established an explainable deep learning system for breast cancer detection through the analysis of 2D grayscale ultrasound images. The MobileNetV2 architecture enabled the model to achieve 90.1% validation accuracy while maintaining real-time processing capabilities and delivering strong performance across all classes, with clear separation between them. The system stands out through its dual capability of medical diagnosis and patient-specific explanation through Grad-CAM visualizations. The system displays important diagnostic areas in ultrasound images, which helps doctors understand AI decisions and build trust in the system for clinical adoption. Our system provides an optimal combination of performance and transparency, and operational speed, which positions it as a suitable solution for practical implementation. The model becomes easily adaptable for different clinical environments because it operates with public datasets and a lightweight architecture. In conclusion the research also presents an innovative AI system that provides human-readable explanations to solve breast cancer diagnosis problems in areas with insufficient radiologic expertise. The system will receive future development to enhance its general applicability and incorporate patient-specific information and obtain necessary regulatory approvals for medical use as well as inclusion for multi-modality in deep learning. Data Availability All data produced in the present study are available upon reasonable request to the authors. https://www.kaggle.com/datasets/sabahesaraki/breast-ultrasound-images-dataset Authors contributions Ghulam H. Abbas developed the AI model, designed the study, concluded the statistical analysis and wrote the manuscript. Shafee Ur Rehman, PhD helped in analyzing the data, sorting references and as the biostatistician verified that all statistical analysis is correct. Ghazal Bargshady, PhD supervised and reviewed the final manuscript for final edits and academic rigor. Ethical Statements No human or animal model was used in this study. all the information collected in this study was obtained from already published results and are available online. Conflict of Interest No conflict of interest Data Availability All the data is available in the manuscript and all data attaining the code and source materials is available on request via the corresponding authors’ email. Funding No funding was available for this research and study. References 1. ↵ Elengoe A . A short review on breast cancer . International Journal of Biotechnology and Biomedicine (IJBB ). 2024 Apr 24; 1 ( 1 ): 1 – 1 . OpenUrl 2. ↵ Barrios CH . Global challenges in breast cancer detection and treatment . The Breast . 2022 Mar 1; 62 : S3 – 6 . OpenUrl 3. ↵ Shahid MS , Imran A . Breast cancer detection using deep learning techniques: challenges and future directions . Multimedia Tools and Applications . 2025 Feb ; 84 ( 6 ): 3257 – 304 . OpenUrl 4. ↵ Rehman SU , Asel U , Abdullah M , Osmonaliev K , Habib A , Shahzad A , Abdillaeva N. The development of predictive biomarkers and immunologic markers for breast cancer: current status and future perspectives . Brazilian Journal of Biology . 2025 May 26; 85 : e292947 . OpenUrl 5. ↵ Schwartz C , Chukwudozie IB , Tejeda S , Vijayasiri G , Abraham I , Remo M , Shah HA , Rojas M , Carillo A , Moreno L , Warnecke RB . Association of population screening for breast cancer risk with use of mammography among women in medically underserved racial and ethnic minority groups . JAMA network open . 2021 Sep 1; 4 ( 9 ): e2123751 -. OpenUrl 6. ↵ Iima M , Le Bihan D . The road to breast cancer screening with diffusion MRI . Frontiers in oncology . 2023 Feb 21; 13 : 993540 . OpenUrl PubMed 7. ↵ Bhatti E , Kaur P . Advances in sensor technologies for breast cancer detection: a comprehensive review of imaging and non-imaging approaches . Discover Artificial Intelligence . 2025 Dec ; 5 ( 1 ): 1 – 6 . OpenUrl 8. ↵ Qaseem A , Lin JS , Mustafa RA , Horwitch CA , Wilt TJ , Clinical Guidelines Committee of the American College of Physicians*. Screening for breast cancer in average-risk women: a guidance statement from the American College of Physicians . Annals of internal medicine . 2019 Apr 16; 170 ( 8 ): 547 – 60 . OpenUrl CrossRef PubMed 9. ↵ Hussain S , Mubeen I , Ullah N , Shah SS , Khan BA , Zahoor M , Ullah R , Khan FA , Sultan MA . Modern diagnostic imaging technique applications and risk factors in the medical field: a review . BioMed research international . 2022 ; 2022 ( 1 ): 5164970 . OpenUrl PubMed 10. ↵ Iacob R , Iacob ER , Stoicescu ER , Ghenciu DM , Cocolea DM , Constantinescu A , Ghenciu LA , Manolescu DL . Evaluating the role of breast ultrasound in early detection of breast cancer in low-and middle-income countries: a comprehensive narrative review . Bioengineering . 2024 Mar 7; 11 ( 3 ): 262 . OpenUrl PubMed 11. ↵ Malherbe K. Imaging of Breast Pathology . In A Mammographers Guide: Radiological and Histopathological Guidelines 2025 Mar 22 (pp. 1 – 52 ). Cham : Springer Nature Switzerland . 12. ↵ Li M , Jiang Y , Zhang Y , Zhu H . Medical image analysis using deep learning algorithms . Frontiers in public health . 2023 Nov 7; 11 : 1273253 . OpenUrl PubMed 13. ↵ Pinto-Coelho L . How artificial intelligence is shaping medical imaging technology: a survey of innovations and applications . Bioengineering . 2023 Dec 18; 10 ( 12 ): 1435 . OpenUrl PubMed 14. ↵ Obuchowicz R , Strzelecki M , Piórkowski A . Clinical applications of artificial intelligence in medical imaging and image processing—A review . Cancers . 2024 May 14; 16 ( 10 ): 1870 . OpenUrl PubMed 15. ↵ Lokaj B , Pugliese MT , Kinkel K , Lovis C , Schmid J . Barriers and facilitators of artificial intelligence conception and implementation for breast imaging diagnosis in clinical practice: a scoping review . European radiology . 2024 Mar ; 34 ( 3 ): 2096 – 109 . OpenUrl PubMed 16. ↵ Durur-Subasi I , Özçelik ŞB. Artificial Intelligence in breast imaging: opportunities, challenges, and legal–ethical considerations . The Eurasian Journal of Medicine . 2023 Dec 1; 55 ( Suppl 1 ): S114 . OpenUrl 17. ↵ Sharafaddini AM , Esfahani KK , Mansouri N . Deep learning approaches to detect breast cancer: a comprehensive review . Multimedia Tools and Applications . 2025 Jun ; 84 ( 21 ): 24079 – 190 . OpenUrl 18. ↵ Al-Dhabyani W , Gomaa M , Khaled H , Fahmy A . Dataset of breast ultrasound images . Data in brief . 2020 Feb 1; 28 : 104863 . OpenUrl PubMed 19. ↵ Gupta K , Chawla N . Analysis of histopathological images for prediction of breast cancer using traditional classifiers with pre-trained CNN . Procedia Computer Science . 2020 Jan 1; 167 : 878 – 89 . OpenUrl 20. ↵ Alom MR , Farid FA , Rahaman MA , Rahman A , Debnath T , Miah AS , Mansor S . An explainable AI-driven deep neural network for accurate breast cancer detection from histopathological and ultrasound images . Scientific Reports . 2025 May 20; 15 ( 1 ): 17531 . OpenUrl PubMed 21. ↵ Gao Z , Tian Y , Lin SC , Lin J. A ct image classification network framework for lung tumors based on pre-trained mobilenetv2 model and transfer learning, and its application and market analysis in the medical field . arXiv preprint arXiv:2501.04996. 2025 Jan 9. 22. ↵ Kandhro IA , Manickam S , Fatima K , Uddin M , Malik U , Naz A , Dandoush A . Performance evaluation of E-VGG19 model: Enhancing real-time skin cancer detection and classification . Heliyon . 2024 May 30; 10 ( 10 ). 23. ↵ Park DJ , Park MW , Lee H , Kim YJ , Kim Y , Park YH . Development of machine learning model for diagnostic disease prediction based on laboratory tests . Scientific reports . 2021 Apr 7; 11 ( 1 ): 7567 . OpenUrl PubMed 24. ↵ Ogundokun RO , Owolawi PA , Tu C . Optimized Deep Feature Learning with Hybrid Ensemble Soft Voting for Early Breast Cancer Histopathological Image Classification. Computers , Materials & Continua . 2025 Sep 1; 84 ( 3 ). 25. Zhao W , Lv J , Zhang H . Breast Cancer Histopathology Image Classification Based on ResNet50 Enhanced with Deep Transfer Learning Technology . InProceedings of the 2024 Guangdong-Hong Kong-Macao Greater Bay Area International Conference on Digital Economy and Artificial Intelligence 2024 Jan 19 (pp. 533 – 538 ). 26. Sharma S , Mehra R . Conventional machine learning and deep learning approach for multi-classification of breast cancer histopathology images—a comparative insight . Journal of digital imaging . 2020 Jun ; 33 ( 3 ): 632 – 54 . OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted November 02, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Advancing Breast Cancer-AI Diagnostics: An Explainable Deep Learning Model Using 2D Grayscale Ultrasound Imaging Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Advancing Breast Cancer-AI Diagnostics: An Explainable Deep Learning Model Using 2D Grayscale Ultrasound Imaging Ghulam Husain Abbas , Shafee Ur Rehman , Ghazal Bargshady medRxiv 2025.10.31.25339239; doi: https://doi.org/10.1101/2025.10.31.25339239 Share This Article: Copy Citation Tools Advancing Breast Cancer-AI Diagnostics: An Explainable Deep Learning Model Using 2D Grayscale Ultrasound Imaging Ghulam Husain Abbas , Shafee Ur Rehman , Ghazal Bargshady medRxiv 2025.10.31.25339239; doi: https://doi.org/10.1101/2025.10.31.25339239 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Health Informatics Subject Areas All Articles Addiction Medicine (572) Allergy and Immunology (864) Anesthesia (302) Cardiovascular Medicine (4448) Dentistry and Oral Medicine (444) Dermatology (383) Emergency Medicine (609) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1515) Epidemiology (15240) Forensic Medicine (30) Gastroenterology (1130) Genetic and Genomic Medicine (6613) Geriatric Medicine (669) Health Economics (1000) Health Informatics (4550) Health Policy (1372) Health Systems and Quality Improvement (1613) Hematology (543) HIV/AIDS (1268) Infectious Diseases (except HIV/AIDS) (15927) Intensive Care and Critical Care Medicine (1105) Medical Education (624) Medical Ethics (147) Nephrology (668) Neurology (6621) Nursing (346) Nutrition (999) Obstetrics and Gynecology (1147) Occupational and Environmental Health (957) Oncology (3342) Ophthalmology (977) Orthopedics (369) Otolaryngology (421) Pain Medicine (436) Palliative Medicine (130) Pathology (665) Pediatrics (1695) Pharmacology and Therapeutics (693) Primary Care Research (714) Psychiatry and Clinical Psychology (5459) Public and Global Health (9247) Radiology and Imaging (2205) Rehabilitation Medicine and Physical Therapy (1371) Respiratory Medicine (1197) Rheumatology (597) Sexual and Reproductive Health (715) Sports Medicine (530) Surgery (714) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a02832f30eb415d5',t:'MTc3OTkxODI0Ng=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();