Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors using Segmentation and Feature Extraction

Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors using Segmentation and Feature Extraction | 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 Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors using Segmentation and Feature Extraction Ali Hamza, Martin Mezl This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3930759/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 Background Breast cancer remains a significant global health challenge, demanding accurate and effective diagnostic methods for timely treatment. Ultrasound imaging stands out as a valuable diagnostic tool for breast cancer due to its affordability, accessibility, and non-ionizing radiation properties. Methods We evaluate the proposed method using a publicly available breast ultrasound images. This paper introduces a novel approach to classifying breast ultrasound images based on segmentation and feature extraction algorithm. The proposed methodology involves several key steps. Firstly, breast ultrasound images undergo preprocessing to enhance image quality and eliminate potential noise. Subsequently, a U-Net + + is applied for the segmentation. A classification model is then trained and validated after extracting features by using Mobilenetv2 and Inceptionv3 of segmented images. This model utilizes modern machine learning and deep learning techniques to distinguish between malignant and benign breast masses. Classification performance is assessed using quantitative metrics, including recall, precision and accuracy. Our results demonstrate improved precision and consistency compared to classification approaches that do not incorporate segmentation and feature extraction. Feature extraction using InceptionV3 and MobileNetV2 showed high accuracy, with MobileNetV2 outperforming InceptionV3 across various classifiers. Results The ANN classifier, when used with MobileNetV2, demonstrated a significant increase in test accuracy (0.9658) compared to InceptionV3 (0.7280). In summary, our findings suggest that the integration of segmentation techniques and feature extraction has the potential to enhance classification algorithms for breast cancer ultrasound images. Conclusion This approach holds promise for supporting radiologists, enhancing diagnostic accuracy, and ultimately improving outcomes for breast cancer patients. In future our focus will be to use comprehensive datasets to validate our methodology. Classification MobilenetV2 InceptionV3 Segmentation Feature extraction Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. Introduction Breast cancer is indeed the most common invasive cancer in women worldwide, as per the international agency for research on cancer, it is projected that there will be approximately 2.7 million new cancer diagnoses and 0.86 million fatalities in the world by 2030 [ 1 ]. Breast cancer is responsible for 19% of the new cases and constitutes 30% of all cancer cases among females. Additionally, it is worth noting that breast cancer incidence rates have been steadily rising by approximately 0.5% annually since the mid-2000s and early detection is crucial for successful treatment [ 2 ]. Ultrasound imaging is a non-invasive and widely available diagnostic tool that can help to evaluate breast masses and identify potential cancerous growths. Compared to other imaging modalities such as X-rays, ultrasound is less expensive and does not expose patients to ionizing radiation. Furthermore, ultrasound is particularly useful for distinguishing between solid masses and fluid-filled cysts [ 3 ], which can help to guide further diagnostic testing and treatment decisions. However, like all medical procedures, the accuracy of ultrasound imaging depends on the skill and experience of the radiologist performing the exam. Therefore, it is important to seek out a qualified and experienced radiologist to ensure the most accurate diagnosis and treatment plan. Computer-aided diagnosis (CAD) systems can help radiologists interpret breast ultrasound images more accurately, and mass segmentation is a critical step in these systems. Accurate segmentation can facilitate better analysis of features related to breast mass shape, which in turn can improve the accuracy of mass classification [ 4 ] [ 5 ] [ 6 ] [ 7 ]. However, automatic segmentation in ultrasound imaging is challenging due to factors such as low image contrast, speckle noise, and variations in breast mass sizes and shapes [ 8 ]. Overall, deep learning-based CAD systems have the potential to significantly improve the accuracy and efficiency of breast mass analysis in ultrasound imaging [ 7 ] [ 9 ] [ 8 ]. However, further research is needed to optimize these methods and to evaluate their clinical utility in real-world scenarios. Deep learning algorithms, such as convolutional neural networks (CNNs), have shown promise in breast mass image analysis [ 10 ]. These methods are data-driven and can automatically learn high-level representations of images to perform segmentation and classification. CNNs have been successfully applied for the detection, segmentation, and classification of breast masses in ultrasound images, and they have demonstrated high accuracy compared to traditional machine learning methods. Various deep learning-based approaches have been proposed for breast mass segmentation and classification in ultrasound imaging [ 4 ] [ 11 ] [ 12 ]. One popular approach is based on convolutional neural networks (CNNs), which can learn to produce pixel-wise segmentation maps directly from input images [ 13 ]. CNNs have been used for breast mass segmentation in 2D and 3D ultrasound images, and they have demonstrated high accuracy compared to traditional segmentation methods [ 10 ] [ 14 ] [ 15 ] [ 16 ]. We implemented two steps algorithm which perform Segmentation using U-Net++ [ 17 ] as a first step. Further analysis has performed to classify the tumour as normal, benign, or malignant. We implemented two approaches to enhance classification results by using deep learning architectures such as InceptionV3 and MobilenetV2 to extract features and implement the classifiers as follows [ 18 ]. Support Vector Machines (SVM) Random Forest Artificial neural network (ANN) K-Nearest Neighbours (KNN) Long Short-Term Memory (LSTM) After completing the tumor classification process, we are well-equipped to provide detailed insights into the specific nature of the identified tumors. This valuable information can assist radiologists in the diagnostic process. In this study our primary focus is to detect tumors accurately and establish a robust framework for the development of patient care. 2. RELATED WORK Dar et al [ 19 ] emphasizes the role of U-Net architecture with Intersection over Union (IoU) is 82.58% in the segmentation, subsequent classification of breast tumors in ultrasound images. The study focuses on refining the standard convolution used in the U-Net architecture to overcome challenges like loss of information and inaccurate boundary localization, which are crucial for effective tumor classification. A Recent study conducted by Pramanik et al [ 20 ] introduces a variant of the U-Net architecture, known as dual branch U-Net (DBU-Net) with an IoU is 74.34%. The dual-branch structure of DBU-Net is designed to enhance feature extraction and improve segmentation accuracy. By improving segmentation, the study indirectly contributes to more accurate tumor classification, as precise segmentation is a critical step in classifying tumors effectively. Moreover, Pun et al [ 21 ]. focuses on residual cross-spatial attention-guided Inception U-Net (RCAIU-Net), integrating residual cross-spatial attention and inception modules within the U-Net framework. This architecture aims to provide more accurate visualizations of breast tumors in ultrasound images with IoU 91%, enhancing classification potential. Zhao et al [ 22 ] also proposes a breast tumor ultrasound image segmentation method using an enhanced U-Net framework with residual blocks and attention mechanisms with IoU is 85%. These enhancements aim to boost the network’s efficiency in tumor recognition and diagnosis, aiding in better classification outcomes. Recently Jabeen et al. [ 23 ] presents a breast cancer classification method using ultrasound images, merging computer vision and deep learning techniques. It involves crucial steps like tumor segmentation, feature extraction, and image preprocessing for enhancing classification accuracy. In another study Zhuang et al. [ 24 ] used fully extracting Image of Interest (IOI) and Region of Interest (ROI) models to extract Areas of Interest (AOIs) from breast ultrasound images. They then used transfer learning combined with SDCB-NET and VGG to classify the extracted AOIs. The results of the study showed that the combination of transfer learning and SDCB-NET/VGG architectures achieved a maximum accuracy of 92.86% in classifying breast ultrasound images. This suggests that the use of deep learning techniques and transfer learning can be effective in improving classification performance. Further, Srikantamurthy et al. [ 25 ] developed a hybrid model using a convolutional neural network (CNN) and long short-term memory recurrent neural network (LSTM-RNN) to classify benign and malignant subtypes of breast cancer. The model, leveraging transfer learning from ImageNet, was evaluated using the BreakHis dataset, comprising 2480 benign and 5429 malignant cancer images. The proposed hybrid CNN-LSTM model achieved a high accuracy of 99% for binary classification and 92.5% for multi-class classification of breast cancer subtypes. Cao et al. [ 26 ] evaluated several CNN models for tumor detection, including Fast Region-based Convolutional Neural Network (FR-CNN), Faster R-CNN, You Only Look Once (YOLO) network, and Single Shot Multibox Detector (SSD). These networks are commonly used for real-time object detection. According to the results the SSD model obtained the highest F1-score of 0.79. Overall, this information suggests that the SSD model is a promising option for tumor detection using CNN models and Yap et al. [ 27 ] compared four CNN models, LeNet, U-Net, and Fully Convolutional Network (FCN-AlexNet), for tumor detection According to the results, FCN-AlexNet model achieved the best performance with an F1-score of 0.92. This suggests that the FCN-AlexNet model is a promising option for tumor detection using CNNs. It is also important to note that all the CNN models outperformed the traditional methods, indicating that CNNs can be a more effective approach for tumor detection compared to manually designed methods. In the study conducted by Podda et al [ 28 ] presents an automated multi-layer process for classifying breast cancer risk from ultrasound images. It involves testing various CNN architectures, combining them into ensembles for improved discrimination, and employing a novel optimization cycle that refines segmentation and classification iteratively. Achieving a Dice coefficient of 82% and a classification accuracy of 91%, the proposed method demonstrates effectiveness, rivaling current leading approaches. Summarizing the recent advancements in the field of Breast Ultrasound images (BUSI) related problems on different ultrasound datasets, it is evident from Table 1 that deep learning architectures have brought significant improvements. The evolution from standard U-Net architectures to enhanced versions like DBU-Net and RCAIU-Net underscores a targeted effort to refine segmentation accuracy and feature extraction, which are pivotal for effective classification. Table 1 Overview of Different Breast Ultrasound Datasets from Literature Review. Reference Methodology Dataset Results (%) Task Dar et al [ 9 ] Efficient-Unet + classification Breast US IoU = 82.58 Recall = 97 Segmentation, Classification Pramanik et al [ 20 ] DBU-Net Breast US IoU = 74.34 Segmentation Pun et al [ 21 ] RCA-IUnet Breast US IoU = 91 Segmentation Zhao et al [ 22 ] Enhanced U-net Breast US IoU = 85 Segmentation Jabeen et al [ 23 ] Modified DarkNet53 Breast US Accuracy = 99 Classification Zhuang et al [ 24 ] SDCB-Net Breast US Accuracy = 92.86 Classification based on feature extraction Srikantamurthy et al [ 25 ] LSTM-RNN Breast US Accuracy = 99(binary) Accuracy = 92.5(multiclass) Classification Cao et al [ 26 ] SSD Breast US F1 score = 79 Detection Yap et al [ 27 ] Four FCN Breast US Accuracy benign = 89.6 Accuracy malignant = 60.6 Classification, Detection Podda et al [ 28 ] CNN Breast US Dice score = 82 Accuracy = 91 Segmentation, Classification 3. PROPOSED METHODOLOGY Our methodology integrates two main components: an initial module leveraging U-Net + + for precise segmentation of breast tumors from ultrasound image followed by a sophisticated classification network that categorizes the segmented tumors. This sequential framework ensures a robust diagnostic pipeline, enhancing the interpretability and accuracy of tumor detection. Further sections will provide an in-depth exploration of the architecture and operational intricacies of both modules. 3.1 U-Net++ Our study involved making enhancements to the U-Net [ 29 ] framework with the aim of enhancing its performance in segmenting breast tumors in US images. We used U-Net++, which is an improved version of the U-Net architecture as shown in Fig. 1 . The architecture of U-Net + + builds upon the original U-Net architecture and introduces nested skip pathways for improved feature aggregation and context modeling. Here is an overview of the architecture of U-Net++. 3.1.1 Encoder pathway U-Net + + starts with an encoder pathway similar to that of U-Net. The encoder is responsible for down sampling the input image and extracting features at different scales. Convolutional layers and max-pooling operations are used in the encoder to reduce the spatial dimensions of the feature maps while increasing the number of channels [ 17 ]. 3.1.2 Nested Skip Pathways The core innovation of U-Net + + is the introduction of nested skip pathways [ 17 ]. These pathways create multiple levels of feature aggregation and context modeling. Instead of a single set of skip connections, U-Net + + establishes nested skip connections at multiple levels [ 15 ]. Each level corresponds to a different scale of feature maps in the encoder. The shortest skip connections directly connect feature maps from the encoder to the decoder at corresponding scales [ 17 ] [ 30 ]. These connections provide high-resolution information to the decoder. Longer skip connections capture features from various levels of the encoder and feed them into the decoder. These connections enable the model to capture context from features at various scales. 3.1.3 Decoder Pathway The decoder pathway in U-Net + + is responsible for up sampling the feature maps and generating the final segmentation mask. Up sampling is achieved through transposed convolution (also known as deconvolution) layers. Feature maps from the encoder are combined with feature maps from the nested skip pathways at each decoder level. Skip connections facilitate the flow of information from both the encoder and the nested skip pathways, enabling the decoder to refine the segmentation mask [ 17 ]. 3.1.4 Final Output The final output of U-Net + + is the segmentation mask, which has the same spatial dimensions as the input image. It represents the predicted segmentation in Fig. 7 . 3.2 Feature extraction using Inceptionv3 and Mobilenetv2 Utilizing the InceptionV3 architecture for feature extraction on segmented breast ultrasound data proved to be a highly effective method for uncovering informative representations within segmented regions of interest (ROIs) [ 31 ]. Following appropriate data preprocessing and resizing of the segmented ROIs to match InceptionV3's input size, the pre-trained InceptionV3 model - with its top classification layer removed - was employed to extract high-dimensional feature vectors from each segmented ROI. These features captured vital patterns and characteristics within the segmented regions, facilitating classification in the later part of the problem. By synergizing the spatial information from segmentation with the semantic information from feature extraction, this approach produced more precise and meaningful insights from breast ultrasound data [ 32 ]. The utilization of the MobileNetV2 model for feature extraction on segmented ultrasound data involves the extraction of informative features from segmented regions of interest (ROIs) within the ultrasound images, using a pre-trained MobileNetV2 model. To achieve this, the segmented ROIs are first resized and align with the model's input size. The MobileNetV2 architecture is then employed to generate feature tensors for each segmented ROI. These feature tensors are designed to capture essential patterns and characteristics within the segmented regions [ 33 ] [ 34 ]. Extracted features are saved and can be utilized for various purposes, such as classification, analysis, or decision-making in medical [ 34 ] image analysis and other applications. This approach enhances the accuracy and significance of insights derived from segmented ultrasound data, contributing to improved diagnostic and analytical capabilities [ 26 ]. 3.3 Classification network The feature extracted vectors generated by the MobileNetv2 and InceptionV3 model is used as input to the classification network. These extracted features are then passed to the classification module [ 35 ] [ 36 ] [ 37 ] [ 38 ], as illustrated in Fig. 3 and Fig. 4 using MobileNetv2 and InceptionV3, respectively. Initially, the features matrix is transformed into a 1D array and fed into a dense layer comprising 120 neurons. This dense layer is followed by a dropout layer with a dropout rate of 0.5 and a batch normalization layer. Finally, the SoftMax activation function is applied in LSTM classifier to produce a probability distribution for the input mask across three classes 3.3.1 CLASSIFICATION MODELS Support Vector Machine (SVM) SVM is a powerful and widely used classification algorithm known for its effectiveness in high-dimensional spaces. We leverage the scikit-learn library to implement SVM with both linear and non-linear kernel functions. The hyperparameters are fine-tuned through cross-validation [ 36 ] [ 35 ]. [ 39 ]. Artificial neural network (ANN) ANNs, specifically designed for non-linear modeling, provide flexibility and adaptability. We construct feedforward neural networks using TensorFlow and Keras, comprising multiple layers with varying units and activation functions. Extensive experimentation is conducted to optimize the architecture [ 40 ] [ 41 ] [ 42 ]. K-nearest Neighbours (KNN) KNN is an instance-based classification method that relies on similarity measures. We apply KNN with varying values of k and explore different distance metrics, tailoring the model to the dataset's characteristics [ 39 ]. Random Forest Random Forest is an ensemble learning method that combines multiple decision trees. We construct a Random Forest classifier with varying tree depths and the number of estimators to harness the strengths of ensemble techniques [ 42 ]. Long short-term memory (LSTM): Incorporating the temporal aspect of the data, we introduce LSTM networks. This model is trained on sequences of feature vectors extracted from segmented BUSI images. LSTM's recurrent architecture is adept at capturing sequential dependencies, making it a valuable addition to our classification arsenal [ 25 ] [ 39 ]. 4. EXPERIMENTAL RESULTS In the following section, we provide a comprehensive overview of the experiments performed using our proposed methodology and discuss the evaluation metrics employed. We also present the results obtained for both segmentation and classification tasks. Afterwards, we conduct a comparative analysis of different classification models, comparing the performance of our proposed model with other state-of-the-art models. In this section, we defined the evaluation metrics used in both the segmentation and classification tasks. For reference and convenience, we have provided the mathematical formulations for these metrics from Eq. 1 to 6. $$\begin{array}{c}{IoU}_{class}=\frac{{TP}_{class}}{{TP}_{class}+{FP}_{class}+{FN}_{class}} \#\left(1\right)\end{array}$$ $$\begin{array}{c}{Dice-score}_{class}=\frac{2{TP}_{class}}{2{TP}_{class}+{FP}_{class}+{FN}_{class}} \#\left(2\right)\end{array}$$ $$\begin{array}{c}{recall}_{class}=\frac{{TP}_{class}}{{TP}_{class}+{FN}_{class}} \#\left(3\right)\end{array}$$ $$\begin{array}{c}{Precision}_{class}=\frac{{TP}_{class}}{{TP}_{class}+{FP}_{class}} \#\left(4\right)\end{array}$$ $$\begin{array}{c}{F1score}_{class}=\frac{2*{recall}_{class}*{Precision}_{class}}{{recall}_{class+}{Precision}_{class}} \#\left(5\right)\end{array}$$ $$\begin{array}{c}Accuracy=\sum \frac{True positives of all classes}{Total number of all classes}\#\left(6\right)\end{array}$$ In the context of segmentation, the Dice score, and IoU serve as similarity metrics, where Dice score quantifies the ratio of twice the area of overlap between the ground truth and predicted mask images to the total number of pixels in both images [ 6 ]. Similarly, IoU measures the overlap between the ground truth and the predicted mask. Both Dice and IoU are useful in measuring image similarity, while accuracy computes the percentage of matching pixels between the predicted mask and ground truth in segmentation tasks [ 43 ]. However, in the case of classification, where there are three distinct classes - benign, malignant, and normal - we use four key metrics: accuracy, precision, recall, and F1-score [ 26 ] [ 44 ]. The formula for the F1-score aligns with that of the dice score, as shown in the form of mathematical equations. Moreover, precision and recall are calculated differently for the classification model. To derive these metrics, we first calculate four essential values: True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN), which are based on the three classes of US images: malignant, benign, and normal. In the following section, we provide a comprehensive overview of the experiments performed using our proposed methodology and discuss the evaluation metrics employed. We also present the results obtained for both segmentation and classification tasks. Afterwards, we conduct a comparative analysis of different classification models, comparing the performance of our proposed model with other state-of-the-art models. 4.1 Dataset A data collection initiative was conducted in 2018 to gather baseline information on women between the ages of 25 and 75 [ 45 ]. The dataset consists of 830 breast ultrasound images of 600 female patients. These images are saved in PNG format and have an average size of 500 x 500 pixels. The dataset is divided into three categories: normal, benign, and malignant. Table 2 provides an overview of the distribution of images across these classes, while Fig. 4 displays sample images from the dataset for visual reference. Additionally, each image has a paired ground truth, presented as a mask image see Fig. 5 . Table 2 Distribution of BUSI dataset Cases Number of images Normal 133 Benign 487 Malignant 210 Total 830 Figure 4: Sample images from provided dataset for all three classes. 4.2 Segmentation The proposed methodology, focusing on the BUSI dataset [ 45 ], addresses the challenge of class imbalance within the dataset. Initially, the dataset is organized into distinct categories representing benign, malignant, and normal breast tissues. To ensure impartial evaluation, a subset of images from each category is randomly reserved in a separate folder for later testing. Augmentation techniques [ 46 ] [ 47 ], including random rotations and horizontal flipping, are then systematically applied to the BUSI dataset in Fig. 6 . Subsequently, a U-Net + + architecture is employed for precise breast tissue segmentation in Fig. 7 [ 17 ]. In the experiment, a standardized image preprocessing approach is applied to ensure consistency. All images, regardless of their original dimensions, are resized to a uniform size of 256x256 pixels. Furthermore, to facilitate efficient and expedited computation using floating-point arithmetic, pixel values in the images are normalized within the range of 0 to 1 by dividing each pixel by 255 [ 34 ] [ 48 ]. The performance assessment of the proposed U-Net + + model involves the evaluation of key metrics such as IoU , Dice-score, and the number of parameters. This evaluation includes a thorough comparison with state-of-the-art methods, all of which were trained and evaluated using identical parameter configurations on the BUSI dataset. The outcomes obtained from applying U-Net + + to the BUSI dataset reveal its superior performance across various dimensions when compared to other models such as U-Net, Efficient and U-Net++ [ 17 ] [ 19 ] [ 29 ] as shown in the Table 3 . Overall, this comprehensive experimental design, encompassing image resizing, normalization, K-fold cross-validation, and the incorporation of the Dice loss function, contributes to robust and thorough evaluation of the proposed methodology's performance. Table 3 Comparison of U-Net + + segmentation results with state-of-the-art models on different BUSI dataset Networks Parameters Dice score IoU U-Net [ 29 ] 9.34M 0.888 0.79 Effecient U-Net [ 19 ] 8.6M 0.904 0.80 U-Net++ 9.7M 0.8961 0.82 4.3 Feature extraction using Mobilenetv2 and Inceptionv3 Accurate and informative feature extraction from segmented BUSI data plays a pivotal role in our comprehensive breast cancer diagnosis framework. In this section, we elaborate on the methodology employed for feature extraction, utilizing InceptionV3 and MobileNetV2 architectures [ 31 ] [ 34 ] concurrently to harness the discriminative characteristics embedded within segmented breast ultrasound images. 4.3.1 Data preprocessing As a preliminary step, all segmented BUSI images undergo standardized preprocessing [ 14 ]. This includes resizing the images to a uniform size of 299x299 pixels for Inceptionv3 whereas for mobilenetv2 input size is 256x256. These preprocessing steps ensure that the segmented images are appropriately formatted and scaled for compatibility with feature extraction models. 4.3.2 Feature Extraction Models: InceptionV3 and MobileNetV2 We employ two state-of-the-art deep learning models for feature extraction: InceptionV3 and MobileNetV2 [ 31 ] [ 34 ]. This model, renowned for its efficacy in image analysis tasks, is loaded without the top classification layers, rendering it suitable for feature extraction [ 49 ] [ 50 ]. InceptionV3 captures intricate and hierarchical features, making it a valuable choice. Each preprocessed segmented BUSI image is fed through the modified InceptionV3 model, yielding a feature vectors that encapsulates high-level and abstract image information [ 51 ]. MobileNetV2 is another pre-trained deep learning architecture renowned for its efficiency and performance. Like InceptionV3, MobileNetV2 is employed without its classification head for the sole purpose of feature extraction. The preprocessed segmented images undergo feature extraction using the MobileNetV2 model. This process results in a feature tensor that captures relevant image features. The feature vectors extracted from both InceptionV3 and MobileNetV2 are systematically stored [ 24 ] [ 31 ] [ 34 ]. These feature datasets collectively serve as the foundation for subsequent classification tasks and further analysis. 4.4 CLASSIFICATION In this section, we delve into the classification phase of our breast cancer diagnosis framework, utilizing the rich feature vectors extracted from segmented BUSI dataset. To comprehensively assess and evaluate the diagnostic performance of our system, we employ a range of classification models, including (SVM), (ANN), (KNN), Random Forest, and (LSTM) networks [ 25 ] [ 35 ] [ 36 ] [ 52 ]. Before delving into the details of each classification model, it is essential to outline the common preprocessing steps applied to our dataset: The entire dataset is divided into training, validation, and test sets using k-fold cross-validation (with k = 5 in our both tasks segmentation and classification). This ensures robust model evaluation and generalization. 4.4.1 Models evaluation and results We assess the performance of each classification model using a range of metrics, including accuracy, precision, recall, and F1-score [ 38 ] [ 51 ] [ 9 ]. The evaluation provides insights into each model's strengths and weaknesses, aiding in the selection of the most suitable approach for breast cancer diagnosis. The classification’s models results obtained using InceptionV3 and MobileNetV2 are thoroughly analyzed and shown in Table 4 and Table 5 . We delve into the implications of the model selection process and offer insights into the models' interpretability, computational efficiency, and diagnostic efficacy using. Performance was evaluated in terms of training, validation, and test accuracies, as well as F1 measure, recall, and precision. The results of the study reveal that the utilization of MobileNetV2 for feature extraction consistently yielded higher test accuracies across all classifiers, in comparison to InceptionV3. Specifically, the ANN classifier demonstrated the most significant variance, with a test accuracy of 0.9658 using MobileNetV2, as opposed to 0.7280 with InceptionV3. The trend was also observed for other performance metrics, including F1 measure, recall, and precision. Furthermore, SVM and Random Forest classifiers exhibited a significant decrease in all performance metrics with InceptionV3, whereas LSTM showed the least performance drop, indicating a lesser dependency on the type of feature extraction used. In conclusion, the study findings suggest that the use of MobileNetV2 leads to superior performance across various classifiers in comparison to InceptionV3. The study's results have significant implications for researchers and practitioners interested in deep learning and its applications to real-world problems. Table 4 Performance of different classifiers using Inceptionv3 Classifier Training accuracy Validation accuracy Test accuracy F1 measure Recall Precision SVM 0.9328 0.7126 0.7376 0.7293 0.7467 0.7598 Random forest 1.00 0.6878 0.6793 0.6575 0.6692 0.7214 KNN 0.8010 0.6672 0.6520 0.6417 0.6410 0.6431 ANN 1.00 0.7018 0.7280 0.7017 0.7684 0.7444 LSTM 0.8492 0.7861 0.7194 0.7434 0.7494 0.7204 Table 5 Performance of different classifiers using Mobilenetv2 Classifier Training accuracy Validation accuracy Test accuracy F1 measure Recall Precision SVM 0.9799 0.9099 0.9501 0.9594 0.9587 0.9603 Random forest 1.00 0.8966 0.9486 0.9481 0.9427 0.9603 KNN 0.9671 0.8949 0.9456 0.9489 0.9579 0.9508 ANN 1.00 0.9204 0.9658 0.9662 0.9660 0.9700 LSTM 0.9850 0.9259 0.9317 0.9334 0.9416 0.9246 Feature extraction using Mobilenetv2 is performed better than InceptionV3, because InceptionV3 has more complex structure and has more hidden layers than MobileNetV2. As results shown, it performed comparatively better in training datasets but when it comes to validation and test accuracy its performance is not adequate. Based on the results shown in Table 3 and Table 4 , we can conclude based on results got from (BUSI) that when the datasets are limited then architectures having simple structure like MobileNetV2 come in handy and valuable for getting desired results. Among all the classifiers used in the experiments of this study ANN performed better than other classifiers. The reason behind the selection of these specific classifiers is that they are widely used in the literature of deep learning related problems [ 53 ] [ 54 ] [ 39 ]. Moreover, this decision also helps to make the study robust in nature. 5. DISCUSSION The present paper proposes a novel breast tumor classification scheme for ultrasound images based on combining feature extraction using InceptionV3 and MobileNetV2 with segmentation information. This study presents a comprehensive examination of breast cancer segmentation-based classification using advanced deep learning techniques. In contrast to previously published works related to deep learning-based breast classification works which only extract the features on breast ultrasound images, our work aims to combine medical knowledge in the network and design the segmentation network to obtain the tumor contour and region to extract more effective features for classification. The utilization of U-Net + + architecture for segmentation and the combined application of InceptionV3 and MobileNetV2 for feature extraction have demonstrated promising results. The results show that our method has a better effect on breast tumor classification than other methods as shown in Table 6 . Compared with other breast classification work Srikanthamurthy et al [ 25 ], including the transfer learning methods based on pre-trained models and shallow network methods, our method obtains better classification performance and the highest precision and accuracy, indicating that our method can classify breast tumor more effectively. The precision and recall metrics are shown in Table 6 , which also demonstrate a notable improvement in accuracy when utilizing deep learning-based classifiers, with the ANN classifier outperforming traditional methods like SVM and Random Forest. This supports the hypothesis that deep learning classifiers can capture complex Jabeen et al, Dar et al [ 23 ] [ 33 ] as shown in Table 1 , non-linear relationships within medical image data more effectively than their traditional counterparts. The performance of every aggregation method, including directly concatenating and directly adding feature representation of the two networks, is better than that of a single branch network, indicating that adding a parallel branch based on segmentation prior is meaningful for breast classification. This confirms that the feature aggregation network based on attention can effectively preserve the effective features and remove the redundant features. Xu et al [ 55 ] proposed RMTL-Net for the simultaneously segmentation and classification of BUSI dataset and achieved accuracy of 91.18% in their model. Compared with other feature aggregation methods, this method has a higher performance, demonstrating that our work can extract more effective and classification-related features when compared to Yaozhong et al [ 56 ] study achieved recall value 91.18%. Arnab et al [ 57 ] used mask images instead of segmented images for that reason the recall parameter outperformed the other state of the art for classification and achieved 96.95 %. Wheres, the segmentation is automatic in our scheme, which can enhance the operability of the system. Our work can provide an effective auxiliary diagnostic tool for the classification of BUSI and help effectively identify tumors in early screening. The study has enlightening implications for breast ultrasound classification and even other types of medical image diagnosis since more effective feature extraction is the basis for further image analysis. While the results are promising, there are limitations that must be acknowledged. The dataset used, while sufficiently varied, is limited in size. Future work should aim to include a larger and more diverse set of images to ensure that the model can generalize across different populations and imaging conditions. Furthermore, the current study does not account for the real-time analysis of ultrasound images, a feature that would be necessary for practical clinical application. Additionally, while our results suggest that MobileNetV2 outperformed InceptionV3 in feature extraction, this may not hold across all datasets or tumor characteristics. The computational efficiency of MobileNetV2, however, makes it an attractive option for real-world applications where resources are limited. The implications of this research are significant, suggesting a path forward where AI can assist radiologists by providing a second opinion and reducing the workload associated with the segmentation and classification of breast tumors. The integration of such AI-based tools into clinical practice could potentially increase the early detection rates of breast cancer, thus improving patient outcomes. Table 6 Comparison of our model with state-of-the-art methods for BUSI datasets References Methodology Accuracy % Precision % Recall % Meng Xu et al [ 55 ] RMTL-Net 91.02 93.34 93.34 Yaozhong et al [ 56 ] Segmentation to Classification Scheme 90.78 - 91.18 Arnab et al [ 57 ] Classification with DCNN(k = 3) using binary masks 94 95.6 96.95 Our model Segmentation-feature extraction-classification 96.58 (test) 97 96 6. CONCLUSION In this research endeavor, we embarked on a journey to enhance the precision and effectiveness of breast cancer diagnosis through the integration of advanced deep learning techniques and classification models. The objective was to provide a robust and reliable framework that assists medical practitioners in early breast cancer detection, improving patient outcomes and treatment strategies. Our comprehensive framework began with the segmentation of breast ultrasound images, where we harnessed the power of U-Net + + to delineate regions of interest. The adoption of segmentation algorithms not only improved the accuracy of breast mass profiling but also paved the way for subsequent feature extraction. Feature extraction, a pivotal step in our methodology, was carried out using both InceptionV3 and MobileNetV2 architectures. These feature-rich vectors encapsulated intricate details and characteristics within the segmented BUSI images. The fusion of features from these models enriched our feature space, enhancing our diagnostic capabilities. The subsequent classification phase was a testament to our commitment to thorough evaluation. Leveraging a diverse ensemble of classifiers, including Support Vector Machines, Artificial Neural Networks, k-Nearest Neighbours, Random Forest, and Long Short-Term Memory networks, we explored multiple avenues for accurate diagnosis. Each classifier brought its own strengths and nuances to the diagnostic task, allowing us to gain a comprehensive understanding of their capabilities. Our extensive experiments, guided by rigorous evaluation metrics, provided valuable insights into the strengths and limitations of each model. Through this analysis, we aimed to empower medical practitioners with a suite of tools that can aid in their decision-making process. In conclusion, our research contributes to the field of breast cancer diagnosis by presenting a holistic framework that amalgamates the prowess of deep learning, image analysis, and classification techniques. While no single model emerged as the panacea, our ensemble of classifiers provides a versatile toolkit that can be tailored to specific diagnostic scenarios and requirements. As we reflect on our journey, we acknowledge that the path to improving breast cancer diagnosis is an ongoing one. Future endeavors may explore additional modalities, extend datasets, and fine-tune model architectures [ 58 ]. Our hope is that this research serves as a steppingstone, inspiring further innovation and collaboration in the pursuit of early breast cancer detection and improved patient care. Declarations data availability The data supports the findings of this study is publicly available. Codes will be provided after acceptance and on request. Ethics declarations Conflict of interest The authors declare that they have no conflict of interest. Ethical standards This study did not use animal experiments. Data availability The data supports the findings of this study is publicly available. Codes will be provided after acceptance and on request. Ethics approval and consent to participate Not applicable Consent for publication Not applicable Funding Not applicable References [Online].Available:https://gco.iarc.fr/tomorrow/en/dataviz/trends?cancers=20&scale=linear&min_zero=1. [Online]. Available: https://www.cancer.org/research/acs-research-news/facts-and-figures-2022.html . B. Malik and J. C. Klock, "Breast cyst fluid analysis correlations with speed of sound using transmission ultrasound," Academic radiology, vol. 26, no. 1, pp. 76-85, 2019. M. Hu, Y. Li and X. Yang, "Breastsam: A study of segment anything model for breast tumor detection in ultrasound images," arXiv preprint arXiv, vol. 23, no. 05, 2023. Q. Huang, F. Zhang and X. Li, "Machine Learning in Ultrasound Computer-Aided Diagnostic Systems:," Biomed Res Int., 2018. A. E.Ilesanmi, U. Chaumrattanakul and S. S. Makhanov, "Methods for the segmentation and classification of breast ultrasound images: a review.," Journal of Ultrasound, vol. 24, pp. 367-382, 2021. J. Chowdary, P. Yogarajah, P. Chaurasia and V. Guruviah, "A Multi-Task Learning Framework for Automated Segmentation and Classification of Breast Tumors From Ultrasound Images," Sage Journals, Ultrasonic Imaging, vol. 44, no. 1, pp. 3-12, 2022. P. Shi, J. Z. A. Rampun and H. Wang, "A hierarchical pipeline for breast boundary segmentation and calcification detection in mammograms.," Computers in biology and medicine, vol. 1, no. 96, pp. 178-188, 2018. R. Karmakar, S. Chatterjee, A. K. Das and A. Mandal, "BCPUML: Breast Cancer Prediction Using Machine Learning Approach—A Performance Analysis.," SN Computer Science,, vol. 4, no. 4, 2023. S. Liu, Y. Wang, X. Yang, B. Lei, L. Liu, S. X. Li, D. Ni and T. Wang, "Deep learning in medical ultrasound analysis: a review," Engineering, vol. 5, no. 2, pp. 261-275, 2019. M. S. K. Inan, F. I. Alam and R. Hasan, "Deep integrated pipeline of segmentation guided classification of breast cancer from ultrasound images," Biomedical Signal Processing and Control, vol. 75, 2022. R. J. G. V. Sloun, R. Cohen and Y. C. Eldar, "Deep Learning in Ultrasound Imaging," Proceedings of the IEEE, vol. 108, no. 1, pp. 11-29, 2020. Y. Xu, Y. Wang, J. Yuan, Q. Cheng, X. Wang and P. L. Carson, "Medical breast ultrasound image segmentation by machine learning," Ultrasonics, vol. 91, pp. 1-9, 2019. D. Perdios, M. Vonlanthen, F. Martinez, M. Arditi and J. P. Thiran, "CNN-based image reconstruction method for ultrafast ultrasound imaging. IEEE Transactions on Ultrasonics," IEEE Transactions on Ultrasound, Ferroelectrics and Frequency Control, vol. 69, no. 4, pp. 1154-1168, 2022. Y. Peng, M. Sonka and D. Z. Chen, "U-Net v2: Rethinking the Skip Connections of U-Net for Medical Image Segmentation.," arXiv preprint arXiv, 2023. N. Vimala and N. Anupama, "Precise segmentation of fetal head in ultrasound images using improved U‐Net model," ETRI Journal, 2023. Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh and J. Liang, "UNet++: A Nested U-Net Architecture for Medical Image Segmentation," in Springer International PublishinG , Granada, 2018. G. Girish, P. Spandana and B. Vasu, "Breast cancer detection using deep learning," Computer Vision and Pattern Recognition, 2023. M. F. Dar and A. Ganivada, "EfficientU-net: a novel deep learning method for breast tumor segmentation and classification in ultrasound images," Neural Processing Letters, vol. 55, pp. 10439-10462, 2023. P. Pramanik, R. Pramanik, F. Schwenker and R. Sarkar, "DBU-Net: Dual branch U-Net for tumor segmentation in breast ultrasound images," Plos one, vol. 18, no. 11, 2023. N. S. Punn and S. Agarwal, "RCA-IUnet: a residual cross-spatial attention-guided inception U-Net model for tumor segmentation in breast ultrasound imaging," Machine Vision and Applications, vol. 33, no. 2, 2022. T. Zhao and H. Dai, "Breast tumor ultrasound image segmentation method based on improved residual u-net network.," Computational Intelligence and Neuroscience, 2022. K. Jabeen, M. A. Khan, M. Alhaisoni, U. Tariq, Y.-D. Zhang, A. Hamza, A. Mickus and R. Damaševičius, "Breast cancer classification from ultrasound images using probability-based optimal deep learning feature fusion," Sensors, vol. 22, no. 3, 2022. Z. Zhuang, W. Ding, S. Zhuang, A. N. J. Raj, J. Wang, W. Zhou and C. Wei, "Tumor classification in automated breast ultrasound (ABUS) based on a modified extracting feature network. ," Computerized Medical Imaging and Graphics, vol. 90, 2021. M. M. Srikantamurthy, V. P. S. Rallabandi, D. B. Dudekula, S. Natarajan and J. Park, "Classification of benign and malignant subtypes of breast cancer histopathology imaging using hybrid CNN-LSTM based transfer learning.," BMC Medical Imaging, vol. 23, no. 1, p. 19, 2023. Z. Cao, L. Duan, G. Yang, T. Yue and Q. Chen, "An experimental study on breast lesion detection and classification from ultrasound images using deep learning architectures.," BMC medical imaging, vol. 19, no. 1, pp. 1-9, 2019. M. H. Yap, M. Goyal, F. M. Osman, R. Martí, E. Denton, A. Juette and R. Zwiggelaar, "Breast ultrasound lesions recognition: end-to-end deep learning approaches," Journal of medical imaging, vol. 6, no. 1, 2018. A. S. Podda, R. Balia, S. Barra, S. Carta, G. Fenu and L. Piano, "Fully-automated deep learning pipeline for segmentation and classification of breast ultrasound images," Journal of computational science, vol. 63, 2022. O. Ronneberger, P. Fischer and T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, vol. 9351, pp. 234-241, 2015. Z. Li, H. Zhang, Z. Li and Z. Ren, "Residual-Attention UNet++: A Nested Residual-Attention U-Net for Medical Image Segmentation," Applied Sciences, vol. 12, no. 14, 2022. P. Zhang, Z. Ma, Y. Zhang, X. Chen and G. Wang, "Improved inception V3 method and its effect on radiologists’ performance of tumor classification with automated breast ultrasound system," Gland Surgery, vol. 10, no. 7, pp. 2232-2245, 2021. T. Fujioka, M. Mori, K. Kubota, J. Oyama, E. Yamaga, Y. Yashima, L. Katsuta, K. Nomura, M. Nara, G. Oda, T. Nakagawa, Y. Kitazume and U. Tateishi, "The utility of deep learning in breast ultrasonic imaging: a review," Diagnostics, vol. 10, no. 12, 2020. N. Mohiuddin, R. A. Dar, M. Rasool and A. Assad, "Breast cancer detection using deep learning: Datasets, methods, and challenges ahead," Computers in biology and medicine, vol. 149, 2022. K. Dong, C. Zhou, Y. Ruan and Y. Li, "MobileNetV2 Model for Image Classification," in 2020 2nd International Conference on Information Technology and Computer Application (ITCA) , Guangzhou, 2021. M.-W. Huang, C.-W. Chen, W.-C. Lin, S.-W. Ke and C.-F. Tsai, "SVM and SVM ensembles in breast cancer prediction," PloS one, vol. 12, no. 1, 2017. S. Guizani, N. Guizani and S. Gharsallaoui, "A Hybrid CNN-SVM Prediction Approach for Breast Cancer Ultrasound Imaging," International Wireless Communications and Mobile Computing (IWCMC), pp. 1174-1578, 2023. A. A. Hossain, J. K. Nisha and F. Johora, "Breast cancer classification from ultrasound images using VGG16 model based transfer learning," International Journal of Image, Graphics and Signal Processing, vol. 13, no. 1, 2023. Z. Khandezamin, M. Naderan and M. J. Rashti, "Detection and classification of breast cancer using logistic regression feature selection and GMDH classifier," Journal of Biomedical Informatics, vol. 111, 2020. M. Bansal, A. Goyal and A. Choudhary, "A comparative analysis of K-Nearest Neighbor, Genetic, Support Vector Machine, Decision Tree, and Long Short Term Memory algorithms in machine learning," Decision Analytics Journal, vol. 3, 2022. H. Işik and S. Arslan, "An Artificial Neural Network classification approach for use the ultrasound in physiotherapy," National Center for Biotechnology Information, vol. 35, no. 6, 2011. J. G. Precious, S. Selvan and R. Avudaiammal, "Classification of abnormalities in breast ultrasound images using ANN, FIS and ANFIS classifier: A comparison," in IOP publishing , coimbatore, 2021. C. V. Narasimhulu, "An automatic feature selection and classification framework for analyzing ultrasound kidney images using dragonfly algorithm and random forest classifier," IET Image Processing, vol. 15, no. 9, pp. 2080-2096, 2021. S. Zhang, M. Liao, J. Wang, Y. Zhu, Y. Zhang, J. Zhang, R. Zheng, L. Lv, D. Zhu, H. Chen and W. Wang, "Fully automatic tumor segmentation of breast ultrasound images with deep learning. ," Journal of Applied Clinical Medical Physics, vol. 24, no. 1, 2023. M. Radak, H. Y. Lafta and H. Fallahi, "Machine learning and deep learning techniques for breast cancer diagnosis and classification: a comprehensive review of medical imaging studies," Journal of Cancer Research and Clinical Oncology, pp. 1-19, 2023. W. Al-Dhabyani, M. Gomma, K. Hussein and A. Fahmy, “Dataset of breast ultrasound images,” Data in brief, vol. 28, 2020. M. Alruily, W. Said, A. M. Mostafa, M. Ezz and M. Elmezain, "Breast Ultrasound Images Augmentation and Segmentation Using GAN with Identity Block and Modified U-Net 3+," Sensors, vol. 23, no. 20, 2023. M. Amiri, R. Brooks, B. Behboodi and H. Rivaz, "Two-stage ultrasound image segmentation using U-Net and test time augmentation," International journal of computer assisted radiology and surgery, vol. 15, no. 6, pp. 981-988, 2020. M. Mendez, S. Sundararaman, L. Probyn and P. N. Tyrrell, "Approaches and Limitations of Machine Learning for Synthetic Ultrasound Generation: A Scoping Review," Journal of Ultrasound in Medicine, vol. 42, no. 12, pp. 2695-2706, 2023. J. Lian, M. Zhang, N. Jiang, W. Bi and X. Dong, "Feature extraction of kidney tissue image based on ultrasound image segmentation," Journal of Healthcare Engineering, 2021. Madhu and R. Kumar, "A hybrid feature extraction technique for content based medical image retrieval using segmentation and clustering techniques.," Multimedia Tools and Applications, vol. 81, no. 6, pp. 8871-8904, 2022. M. F. Mridha, M. A. Hamid, M. M. Monowar, A. J. Keya, A. Q. Ohi, M. R. Islam and J.-M. Kim, "A comprehensive survey on deep-learning-based breast cancer diagnosis," Cancers, vol. 13, no. 23, 2021. M. Botlagunta, M. D. Botlagunta, M. B. Myneni, D. Lakshmi, A. Nayyar, J. S. Gullapalli and M. A. Shah, "Classification and diagnostic prediction of breast cancer metastasis on clinical data using machine learning algorithms," Scientific Reports, vol. 13, no. 1, 2023. N. Saffari, H. A. Rashwan, M. Abdel-Nasser, V. K. Singh, M. Arenas, E. Mangina, B. Herrera and D. Puig, "Fully automated breast density segmentation and classification using deep learning.," Diagnostics, vol. 10, no. 11, 2020. D. R. Sarvamangala and R. V. Kulkarni, "Convolutional neural networks in medical image understanding: a survey," Evolutionary intelligence,, vol. 15, no. 1, pp. 1-22, 2022. M. Xu, K. Huang and X. Qi, "A Regional-Attentive Multi-Task Learning Framework for Breast Ultrasound Image Segmentation and Classification," IEEE Access, vol. 11, pp. 5377-5392, 2023. Y. Luo, Q. Huang and X. Li, "Segmentation information with attention integration for classification of breast tumor in ultrasound image," Pattern Recognition, vol. 124, 2022. Arnab Kumar Mishra, S. B. Pinki Roy and S. K. Das, "Achieving highly efficient breast ultrasound tumor classification with deep convolutional neural networks," International Journal of information technology, vol. 12, pp. 3311-3320, 2022. H. Afrin, N. B. Larson, M. Fatemi and A. Alizad, "Deep Learning in Different Ultrasound Methods for Breast Cancer, from Diagnosis to Prognosis : Current Trends, Challenges, and an Analysis," Cancers, vol. 15, no. 12, pp. 31-39, 2023. Additional Declarations No competing interests reported. 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 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-3930759","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":271377293,"identity":"3c7db163-842b-485a-ad70-e024575f0600","order_by":0,"name":"Ali Hamza","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvElEQVRIiWNgGAWjYDACdsYGZgijgVgtzDAtPAcYGA4QpwWMgEAigUgt/M3MDcwFNdvs+We+MZP+wGBnT1CLxGGgw2Ycu80scTvHTOIAQzIzYWtAWnjYbrMx3E5LA2o5wEZQhzxYy7/bPPI3j4G18BDUYgDSwtt2W8LgBvMxkBYJgloMgVoO8/bdNjA8k3zY4oxBsgFBLXLH2x8+5vl2217u+MHGGxUVRIQYCBxAcidRGkbBKBgFo2AUEAIAM+Q2qak0k/gAAAAASUVORK5CYII=","orcid":"","institution":"Brno University of Technology","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Ali","middleName":"","lastName":"Hamza","suffix":""},{"id":271377294,"identity":"d41dba9b-0c31-42d7-a602-ba7cf1f38647","order_by":1,"name":"Martin Mezl","email":"","orcid":"","institution":"Brno University of Technology","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Martin","middleName":"","lastName":"Mezl","suffix":""}],"badges":[],"createdAt":"2024-02-05 11:29:44","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3930759/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3930759/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":50925044,"identity":"c5b77c4d-0054-4e5c-8e52-98b6841b43d9","added_by":"auto","created_at":"2024-02-09 17:05:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":191133,"visible":true,"origin":"","legend":"\u003cp\u003eFlow of Segmentation Task Using U-Net++\u003c/p\u003e","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/ab79e260a051631887b37b4d.png"},{"id":50925043,"identity":"32283447-367e-48c9-b67e-8d91293e0c11","added_by":"auto","created_at":"2024-02-09 17:05:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":360663,"visible":true,"origin":"","legend":"\u003cp\u003eDetailed Workflow of Proposed Methodology Using MobileNetv2\u003c/p\u003e","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/3676e235ea5437954c26400a.png"},{"id":50925045,"identity":"9c2561a7-2920-4fc0-8e32-e4aee8a493ad","added_by":"auto","created_at":"2024-02-09 17:05:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":353184,"visible":true,"origin":"","legend":"\u003cp\u003eDetailed Workflow of Methodology Using InceptionV3\u003c/p\u003e","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/c0bf0422c4fc2d313ad81da6.png"},{"id":50925047,"identity":"960a4d47-0278-4681-a44d-a44fad96bdd8","added_by":"auto","created_at":"2024-02-09 17:05:56","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":123622,"visible":true,"origin":"","legend":"\u003cp\u003eSample images from provided dataset for all three classes.\u003c/p\u003e","description":"","filename":"figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/06225a9445ed6d20d4575619.png"},{"id":50925046,"identity":"d7dfc287-cd73-4c38-9f19-337ac84a5369","added_by":"auto","created_at":"2024-02-09 17:05:56","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":106533,"visible":true,"origin":"","legend":"\u003cp\u003eReal Images with respective mask images in original BUSI\u003c/p\u003e","description":"","filename":"figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/7692a87ef36ba62e3ba9ef89.png"},{"id":50925049,"identity":"b16d773c-bf59-42ba-b729-a4f8d0e34a64","added_by":"auto","created_at":"2024-02-09 17:05:56","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":213015,"visible":true,"origin":"","legend":"\u003cp\u003eExample of image augmentation for various images\u003c/p\u003e","description":"","filename":"figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/5251738bb4c88f9221fed06e.png"},{"id":50927119,"identity":"70279c0f-568d-43b8-83eb-ce3fb82dce99","added_by":"auto","created_at":"2024-02-09 17:13:56","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":484667,"visible":true,"origin":"","legend":"\u003cp\u003eExample of segmentation results for various input images. Left column are original image data, middle column is Ground truth reference and right column is a result of segmentation process\u003c/p\u003e","description":"","filename":"figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/1dbdfba63f11cbdfa9038dd0.png"},{"id":53115542,"identity":"9c662eaa-6fb7-47e4-a9bf-2347db0b2ecc","added_by":"auto","created_at":"2024-03-20 19:11:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2229044,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3930759/v1/1a92c52f-60da-4f3a-b040-3d0ce89f71d9.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors using Segmentation and Feature Extraction","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBreast cancer is indeed the most common invasive cancer in women worldwide, as per the international agency for research on cancer, it is projected that there will be approximately 2.7\u0026nbsp;million new cancer diagnoses and 0.86\u0026nbsp;million fatalities in the world by 2030 [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Breast cancer is responsible for 19% of the new cases and constitutes 30% of all cancer cases among females. Additionally, it is worth noting that breast cancer incidence rates have been steadily rising by approximately 0.5% annually since the mid-2000s and early detection is crucial for successful treatment [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Ultrasound imaging is a non-invasive and widely available diagnostic tool that can help to evaluate breast masses and identify potential cancerous growths. Compared to other imaging modalities such as X-rays, ultrasound is less expensive and does not expose patients to ionizing radiation. Furthermore, ultrasound is particularly useful for distinguishing between solid masses and fluid-filled cysts [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], which can help to guide further diagnostic testing and treatment decisions. However, like all medical procedures, the accuracy of ultrasound imaging depends on the skill and experience of the radiologist performing the exam. Therefore, it is important to seek out a qualified and experienced radiologist to ensure the most accurate diagnosis and treatment plan.\u003c/p\u003e \u003cp\u003eComputer-aided diagnosis (CAD) systems can help radiologists interpret breast ultrasound images more accurately, and mass segmentation is a critical step in these systems. Accurate segmentation can facilitate better analysis of features related to breast mass shape, which in turn can improve the accuracy of mass classification [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, automatic segmentation in ultrasound imaging is challenging due to factors such as low image contrast, speckle noise, and variations in breast mass sizes and shapes [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Overall, deep learning-based CAD systems have the potential to significantly improve the accuracy and efficiency of breast mass analysis in ultrasound imaging [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. However, further research is needed to optimize these methods and to evaluate their clinical utility in real-world scenarios.\u003c/p\u003e \u003cp\u003eDeep learning algorithms, such as convolutional neural networks (CNNs), have shown promise in breast mass image analysis [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. These methods are data-driven and can automatically learn high-level representations of images to perform segmentation and classification. CNNs have been successfully applied for the detection, segmentation, and classification of breast masses in ultrasound images, and they have demonstrated high accuracy compared to traditional machine learning methods. Various deep learning-based approaches have been proposed for breast mass segmentation and classification in ultrasound imaging [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. One popular approach is based on convolutional neural networks (CNNs), which can learn to produce pixel-wise segmentation maps directly from input images [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. CNNs have been used for breast mass segmentation in 2D and 3D ultrasound images, and they have demonstrated high accuracy compared to traditional segmentation methods [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWe implemented two steps algorithm which perform Segmentation using U-Net++ [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] as a first step. Further analysis has performed to classify the tumour as normal, benign, or malignant. We implemented two approaches to enhance classification results by using deep learning architectures such as InceptionV3 and MobilenetV2 to extract features and implement the classifiers as follows [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eSupport Vector Machines (SVM)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eArtificial neural network (ANN)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eK-Nearest Neighbours (KNN)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eLong Short-Term Memory (LSTM)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eAfter completing the tumor classification process, we are well-equipped to provide detailed insights into the specific nature of the identified tumors. This valuable information can assist radiologists in the diagnostic process. In this study our primary focus is to detect tumors accurately and establish a robust framework for the development of patient care.\u003c/p\u003e"},{"header":"2.\tRELATED WORK","content":"\u003cp\u003eDar et al [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] emphasizes the role of U-Net architecture with Intersection over Union (IoU) is 82.58% in the segmentation, subsequent classification of breast tumors in ultrasound images. The study focuses on refining the standard convolution used in the U-Net architecture to overcome challenges like loss of information and inaccurate boundary localization, which are crucial for effective tumor classification. A Recent study conducted by Pramanik et al [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] introduces a variant of the U-Net architecture, known as dual branch U-Net (DBU-Net) with an \u003cem\u003eIoU\u003c/em\u003e is 74.34%. The dual-branch structure of DBU-Net is designed to enhance feature extraction and improve segmentation accuracy. By improving segmentation, the study indirectly contributes to more accurate tumor classification, as precise segmentation is a critical step in classifying tumors effectively. Moreover, Pun et al [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. focuses on residual cross-spatial attention-guided Inception U-Net (RCAIU-Net), integrating residual cross-spatial attention and inception modules within the U-Net framework. This architecture aims to provide more accurate visualizations of breast tumors in ultrasound images with \u003cem\u003eIoU\u003c/em\u003e 91%, enhancing classification potential. Zhao et al [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] also proposes a breast tumor ultrasound image segmentation method using an enhanced U-Net framework with residual blocks and attention mechanisms with \u003cem\u003eIoU\u003c/em\u003e is 85%. These enhancements aim to boost the network\u0026rsquo;s efficiency in tumor recognition and diagnosis, aiding in better classification outcomes.\u003c/p\u003e \u003cp\u003eRecently Jabeen et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] presents a breast cancer classification method using ultrasound images, merging computer vision and deep learning techniques. It involves crucial steps like tumor segmentation, feature extraction, and image preprocessing for enhancing classification accuracy. In another study Zhuang et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] used fully extracting Image of Interest (IOI) and Region of Interest (ROI) models to extract Areas of Interest (AOIs) from breast ultrasound images. They then used transfer learning combined with SDCB-NET and VGG to classify the extracted AOIs. The results of the study showed that the combination of transfer learning and SDCB-NET/VGG architectures achieved a maximum accuracy of 92.86% in classifying breast ultrasound images. This suggests that the use of deep learning techniques and transfer learning can be effective in improving classification performance. Further, Srikantamurthy et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] developed a hybrid model using a convolutional neural network (CNN) and long short-term memory recurrent neural network (LSTM-RNN) to classify benign and malignant subtypes of breast cancer. The model, leveraging transfer learning from ImageNet, was evaluated using the BreakHis dataset, comprising 2480 benign and 5429 malignant cancer images. The proposed hybrid CNN-LSTM model achieved a high accuracy of 99% for binary classification and 92.5% for multi-class classification of breast cancer subtypes.\u003c/p\u003e \u003cp\u003eCao et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] evaluated several CNN models for tumor detection, including Fast Region-based Convolutional Neural Network (FR-CNN), Faster R-CNN, You Only Look Once (YOLO) network, and Single Shot Multibox Detector (SSD). These networks are commonly used for real-time object detection. According to the results the SSD model obtained the highest F1-score of 0.79. Overall, this information suggests that the SSD model is a promising option for tumor detection using CNN models and Yap et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] compared four CNN models, LeNet, U-Net, and Fully Convolutional Network (FCN-AlexNet), for tumor detection According to the results, FCN-AlexNet model achieved the best performance with an F1-score of 0.92. This suggests that the FCN-AlexNet model is a promising option for tumor detection using CNNs. It is also important to note that all the CNN models outperformed the traditional methods, indicating that CNNs can be a more effective approach for tumor detection compared to manually designed methods. In the study conducted by Podda et al [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] presents an automated multi-layer process for classifying breast cancer risk from ultrasound images. It involves testing various CNN architectures, combining them into ensembles for improved discrimination, and employing a novel optimization cycle that refines segmentation and classification iteratively. Achieving a Dice coefficient of 82% and a classification accuracy of 91%, the proposed method demonstrates effectiveness, rivaling current leading approaches.\u003c/p\u003e \u003cp\u003eSummarizing the recent advancements in the field of Breast Ultrasound images (BUSI) related problems on different ultrasound datasets, it is evident from Table\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e that deep learning architectures have brought significant improvements. The evolution from standard U-Net architectures to enhanced versions like DBU-Net and RCAIU-Net underscores a targeted effort to refine segmentation accuracy and feature extraction, which are pivotal for effective classification.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\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\u003eOverview of Different Breast Ultrasound Datasets from Literature Review.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReference\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMethodology\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eResults (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTask\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDar et al [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficient-Unet\u0026thinsp;+\u0026thinsp;classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eIoU\u003c/em\u003e\u0026thinsp;=\u0026thinsp;82.58\u003c/p\u003e \u003cp\u003eRecall\u0026thinsp;=\u0026thinsp;97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSegmentation,\u003c/p\u003e \u003cp\u003eClassification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePramanik et al [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDBU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eIoU\u003c/em\u003e\u0026thinsp;=\u0026thinsp;74.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSegmentation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePun et al [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRCA-IUnet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eIoU\u003c/em\u003e\u0026thinsp;=\u0026thinsp;91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSegmentation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZhao et al [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnhanced U-net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eIoU\u0026thinsp;=\u0026thinsp;85\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSegmentation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJabeen et al [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModified DarkNet53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u0026thinsp;=\u0026thinsp;99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eClassification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZhuang et al [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSDCB-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u0026thinsp;=\u0026thinsp;92.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eClassification based on feature extraction\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSrikantamurthy et al [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLSTM-RNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u0026thinsp;=\u0026thinsp;99(binary)\u003c/p\u003e \u003cp\u003eAccuracy\u0026thinsp;=\u0026thinsp;92.5(multiclass)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eClassification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCao et al [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSSD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eF1 score\u0026thinsp;=\u0026thinsp;79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDetection\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYap et al [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFour FCN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy benign\u0026thinsp;=\u0026thinsp;89.6\u003c/p\u003e \u003cp\u003eAccuracy malignant\u0026thinsp;=\u0026thinsp;60.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eClassification,\u003c/p\u003e \u003cp\u003eDetection\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePodda\u0026nbsp;et al [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBreast US\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDice score\u0026thinsp;=\u0026thinsp;82\u003c/p\u003e \u003cp\u003eAccuracy\u0026thinsp;=\u0026thinsp;91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSegmentation,\u003c/p\u003e \u003cp\u003eClassification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"3.\tPROPOSED METHODOLOGY","content":"\u003cp\u003eOur methodology integrates two main components: an initial module leveraging U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;for precise segmentation of breast tumors from ultrasound image followed by a sophisticated classification network that categorizes the segmented tumors. This sequential framework ensures a robust diagnostic pipeline, enhancing the interpretability and accuracy of tumor detection. Further sections will provide an in-depth exploration of the architecture and operational intricacies of both modules.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 U-Net++\u003c/h2\u003e \u003cp\u003eOur study involved making enhancements to the U-Net [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] framework with the aim of enhancing its performance in segmenting breast tumors in US images. We used U-Net++, which is an improved version of the U-Net architecture as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe architecture of U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;builds upon the original U-Net architecture and introduces nested skip pathways for improved feature aggregation and context modeling. Here is an overview of the architecture of U-Net++.\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 Encoder pathway\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eU-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;starts with an encoder pathway similar to that of U-Net. The encoder is responsible for down sampling the input image and extracting features at different scales. Convolutional layers and max-pooling operations are used in the encoder to reduce the spatial dimensions of the feature maps while increasing the number of channels [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Nested Skip Pathways\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe core innovation of U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;is the introduction of nested skip pathways [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. These pathways create multiple levels of feature aggregation and context modeling. Instead of a single set of skip connections, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;establishes nested skip connections at multiple levels [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Each level corresponds to a different scale of feature maps in the encoder. The shortest skip connections directly connect feature maps from the encoder to the decoder at corresponding scales [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. These connections provide high-resolution information to the decoder. Longer skip connections capture features from various levels of the encoder and feed them into the decoder. These connections enable the model to capture context from features at various scales.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e3.1.3 Decoder Pathway\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe decoder pathway in U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;is responsible for up sampling the feature maps and generating the final segmentation mask. Up sampling is achieved through transposed convolution (also known as deconvolution) layers. Feature maps from the encoder are combined with feature maps from the nested skip pathways at each decoder level. Skip connections facilitate the flow of information from both the encoder and the nested skip pathways, enabling the decoder to refine the segmentation mask [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e3.1.4 Final Output\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe final output of U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;is the segmentation mask, which has the same spatial dimensions as the input image. It represents the predicted segmentation in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Feature extraction using Inceptionv3 and Mobilenetv2\u003c/h2\u003e \u003cp\u003eUtilizing the InceptionV3 architecture for feature extraction on segmented breast ultrasound data proved to be a highly effective method for uncovering informative representations within segmented regions of interest (ROIs) [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Following appropriate data preprocessing and resizing of the segmented ROIs to match InceptionV3's input size, the pre-trained InceptionV3 model - with its top classification layer removed - was employed to extract high-dimensional feature vectors from each segmented ROI. These features captured vital patterns and characteristics within the segmented regions, facilitating classification in the later part of the problem. By synergizing the spatial information from segmentation with the semantic information from feature extraction, this approach produced more precise and meaningful insights from breast ultrasound data [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe utilization of the MobileNetV2 model for feature extraction on segmented ultrasound data involves the extraction of informative features from segmented regions of interest (ROIs) within the ultrasound images, using a pre-trained MobileNetV2 model. To achieve this, the segmented ROIs are first resized and align with the model's input size. The MobileNetV2 architecture is then employed to generate feature tensors for each segmented ROI. These feature tensors are designed to capture essential patterns and characteristics within the segmented regions [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Extracted features are saved and can be utilized for various purposes, such as classification, analysis, or decision-making in medical [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] image analysis and other applications. This approach enhances the accuracy and significance of insights derived from segmented ultrasound data, contributing to improved diagnostic and analytical capabilities [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Classification network\u003c/h2\u003e \u003cp\u003eThe feature extracted vectors generated by the MobileNetv2 and InceptionV3 model is used as input to the classification network. These extracted features are then passed to the classification module [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e] [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;4 using MobileNetv2 and InceptionV3, respectively. Initially, the features matrix is transformed into a 1D array and fed into a dense layer comprising 120 neurons. This dense layer is followed by a dropout layer with a dropout rate of 0.5 and a batch normalization layer. Finally, the SoftMax activation function is applied in LSTM classifier to produce a probability distribution for the input mask across three classes\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1 CLASSIFICATION MODELS\u003c/h2\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eSupport Vector Machine (SVM)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSVM is a powerful and widely used classification algorithm known for its effectiveness in high-dimensional spaces. We leverage the scikit-learn library to implement SVM with both linear and non-linear kernel functions. The hyperparameters are fine-tuned through cross-validation [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eArtificial neural network (ANN)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eANNs, specifically designed for non-linear modeling, provide flexibility and adaptability. We construct feedforward neural networks using TensorFlow and Keras, comprising multiple layers with varying units and activation functions. Extensive experimentation is conducted to optimize the architecture [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e] [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eK-nearest Neighbours (KNN)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eKNN is an instance-based classification method that relies on similarity measures. We apply KNN with varying values of k and explore different distance metrics, tailoring the model to the dataset's characteristics [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eRandom Forest is an ensemble learning method that combines multiple decision trees. We construct a Random Forest classifier with varying tree depths and the number of estimators to harness the strengths of ensemble techniques [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eLong short-term memory (LSTM):\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIncorporating the temporal aspect of the data, we introduce LSTM networks. This model is trained on sequences of feature vectors extracted from segmented BUSI images. LSTM's recurrent architecture is adept at capturing sequential dependencies, making it a valuable addition to our classification arsenal [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4.\tEXPERIMENTAL RESULTS","content":"\u003cp\u003eIn the following section, we provide a comprehensive overview of the experiments performed using our proposed methodology and discuss the evaluation metrics employed. We also present the results obtained for both segmentation and classification tasks. Afterwards, we conduct a comparative analysis of different classification models, comparing the performance of our proposed model with other state-of-the-art models.\u003c/p\u003e \u003cp\u003eIn this section, we defined the evaluation metrics used in both the segmentation and classification tasks. For reference and convenience, we have provided the mathematical formulations for these metrics from Eq.\u0026nbsp;1 to 6.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{IoU}_{class}=\\frac{{TP}_{class}}{{TP}_{class}+{FP}_{class}+{FN}_{class}} \\#\\left(1\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{Dice-score}_{class}=\\frac{2{TP}_{class}}{2{TP}_{class}+{FP}_{class}+{FN}_{class}} \\#\\left(2\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{recall}_{class}=\\frac{{TP}_{class}}{{TP}_{class}+{FN}_{class}} \\#\\left(3\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{Precision}_{class}=\\frac{{TP}_{class}}{{TP}_{class}+{FP}_{class}} \\#\\left(4\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}{F1score}_{class}=\\frac{2*{recall}_{class}*{Precision}_{class}}{{recall}_{class+}{Precision}_{class}} \\#\\left(5\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\begin{array}{c}Accuracy=\\sum \\frac{True positives of all classes}{Total number of all classes}\\#\\left(6\\right)\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn the context of segmentation, the Dice score, and \u003cem\u003eIoU\u003c/em\u003e serve as similarity metrics, where Dice score quantifies the ratio of twice the area of overlap between the ground truth and predicted mask images to the total number of pixels in both images [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Similarly, \u003cem\u003eIoU\u003c/em\u003e measures the overlap between the ground truth and the predicted mask. Both Dice and \u003cem\u003eIoU\u003c/em\u003e are useful in measuring image similarity, while accuracy computes the percentage of matching pixels between the predicted mask and ground truth in segmentation tasks [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]. However, in the case of classification, where there are three distinct classes - benign, malignant, and normal - we use four key metrics: accuracy, precision, recall, and F1-score [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. The formula for the F1-score aligns with that of the dice score, as shown in the form of mathematical equations. Moreover, precision and recall are calculated differently for the classification model. To derive these metrics, we first calculate four essential values: True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN), which are based on the three classes of US images: malignant, benign, and normal.\u003c/p\u003e \u003cp\u003eIn the following section, we provide a comprehensive overview of the experiments performed using our proposed methodology and discuss the evaluation metrics employed. We also present the results obtained for both segmentation and classification tasks. Afterwards, we conduct a comparative analysis of different classification models, comparing the performance of our proposed model with other state-of-the-art models.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Dataset\u003c/h2\u003e \u003cp\u003eA data collection initiative was conducted in 2018 to gather baseline information on women between the ages of 25 and 75 [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. The dataset consists of 830 breast ultrasound images of 600 female patients. These images are saved in PNG format and have an average size of 500 x 500 pixels. The dataset is divided into three categories: normal, benign, and malignant. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e provides an overview of the distribution of images across these classes, while Fig.\u0026nbsp;4 displays sample images from the dataset for visual reference. Additionally, each image has a paired ground truth, presented as a mask image see Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\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\u003eDistribution of BUSI dataset\u003cdiv description=\"An ultrasound of a person's bodyDescription automatically generated\" class=\"Drawing\" id=\"41\" name=\"Picture 41\"\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCases\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of images\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e133\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBenign\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e487\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMalignant\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e210\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e830\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFigure 4: Sample images from provided dataset for all three classes.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Segmentation\u003c/h2\u003e \u003cp\u003eThe proposed methodology, focusing on the BUSI dataset [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e], addresses the challenge of class imbalance within the dataset. Initially, the dataset is organized into distinct categories representing benign, malignant, and normal breast tissues. To ensure impartial evaluation, a subset of images from each category is randomly reserved in a separate folder for later testing. Augmentation techniques [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e] [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e], including random rotations and horizontal flipping, are then systematically applied to the BUSI dataset in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Subsequently, a U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;architecture is employed for precise breast tissue segmentation in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e7\u003c/span\u003e [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the experiment, a standardized image preprocessing approach is applied to ensure consistency. All images, regardless of their original dimensions, are resized to a uniform size of 256x256 pixels. Furthermore, to facilitate efficient and expedited computation using floating-point arithmetic, pixel values in the images are normalized within the range of 0 to 1 by dividing each pixel by 255 [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe performance assessment of the proposed U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;model involves the evaluation of key metrics such as \u003cem\u003eIoU\u003c/em\u003e, Dice-score, and the number of parameters. This evaluation includes a thorough comparison with state-of-the-art methods, all of which were trained and evaluated using identical parameter configurations on the BUSI dataset. The outcomes obtained from applying U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;to the BUSI dataset reveal its superior performance across various dimensions when compared to other models such as U-Net, Efficient and U-Net++ [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] as shown in the Table \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eOverall, this comprehensive experimental design, encompassing image resizing, normalization, K-fold cross-validation, and the incorporation of the Dice loss function, contributes to robust and thorough evaluation of the proposed methodology's performance.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\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\u003eComparison of U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;segmentation results with state-of-the-art models on different BUSI dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNetworks\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eParameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDice score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eIoU\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eU-Net [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.34M\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.888\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEffecient U-Net [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.6M\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.904\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eU-Net++\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.7M\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8961\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.82\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Feature extraction using Mobilenetv2 and Inceptionv3\u003c/h2\u003e \u003cp\u003eAccurate and informative feature extraction from segmented BUSI data plays a pivotal role in our comprehensive breast cancer diagnosis framework. In this section, we elaborate on the methodology employed for feature extraction, utilizing InceptionV3 and MobileNetV2 architectures [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] concurrently to harness the discriminative characteristics embedded within segmented breast ultrasound images.\u003c/p\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e4.3.1 Data preprocessing\u003c/h2\u003e \u003cp\u003eAs a preliminary step, all segmented BUSI images undergo standardized preprocessing [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. This includes resizing the images to a uniform size of 299x299 pixels for Inceptionv3 whereas for mobilenetv2 input size is 256x256. These preprocessing steps ensure that the segmented images are appropriately formatted and scaled for compatibility with feature extraction models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e4.3.2 Feature Extraction Models: InceptionV3 and MobileNetV2\u003c/h2\u003e \u003cp\u003eWe employ two state-of-the-art deep learning models for feature extraction: InceptionV3 and MobileNetV2 [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. This model, renowned for its efficacy in image analysis tasks, is loaded without the top classification layers, rendering it suitable for feature extraction [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e] [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. InceptionV3 captures intricate and hierarchical features, making it a valuable choice. Each preprocessed segmented BUSI image is fed through the modified InceptionV3 model, yielding a feature vectors that encapsulates high-level and abstract image information [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMobileNetV2 is another pre-trained deep learning architecture renowned for its efficiency and performance. Like InceptionV3, MobileNetV2 is employed without its classification head for the sole purpose of feature extraction. The preprocessed segmented images undergo feature extraction using the MobileNetV2 model. This process results in a feature tensor that captures relevant image features.\u003c/p\u003e \u003cp\u003eThe feature vectors extracted from both InceptionV3 and MobileNetV2 are systematically stored [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. These feature datasets collectively serve as the foundation for subsequent classification tasks and further analysis.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.4 CLASSIFICATION\u003c/h2\u003e \u003cp\u003eIn this section, we delve into the classification phase of our breast cancer diagnosis framework, utilizing the rich feature vectors extracted from segmented BUSI dataset. To comprehensively assess and evaluate the diagnostic performance of our system, we employ a range of classification models, including (SVM), (ANN), (KNN), Random Forest, and (LSTM) networks [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBefore delving into the details of each classification model, it is essential to outline the common preprocessing steps applied to our dataset: The entire dataset is divided into training, validation, and test sets using k-fold cross-validation (with k\u0026thinsp;=\u0026thinsp;5 in our both tasks segmentation and classification). This ensures robust model evaluation and generalization.\u003c/p\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e4.4.1 Models evaluation and results\u003c/h2\u003e \u003cp\u003eWe assess the performance of each classification model using a range of metrics, including accuracy, precision, recall, and F1-score [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e] [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e] [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The evaluation provides insights into each model's strengths and weaknesses, aiding in the selection of the most suitable approach for breast cancer diagnosis. The classification\u0026rsquo;s models results obtained using InceptionV3 and MobileNetV2 are thoroughly analyzed and shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. We delve into the implications of the model selection process and offer insights into the models' interpretability, computational efficiency, and diagnostic efficacy using. Performance was evaluated in terms of training, validation, and test accuracies, as well as F1 measure, recall, and precision. The results of the study reveal that the utilization of MobileNetV2 for feature extraction consistently yielded higher test accuracies across all classifiers, in comparison to InceptionV3. Specifically, the ANN classifier demonstrated the most significant variance, with a test accuracy of 0.9658 using MobileNetV2, as opposed to 0.7280 with InceptionV3. The trend was also observed for other performance metrics, including F1 measure, recall, and precision. Furthermore, SVM and Random Forest classifiers exhibited a significant decrease in all performance metrics with InceptionV3, whereas LSTM showed the least performance drop, indicating a lesser dependency on the type of feature extraction used. In conclusion, the study findings suggest that the use of MobileNetV2 leads to superior performance across various classifiers in comparison to InceptionV3. The study's results have significant implications for researchers and practitioners interested in deep learning and its applications to real-world problems.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\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\u003ePerformance of different classifiers using Inceptionv3\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClassifier\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTest accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1 measure\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9328\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7126\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.7376\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.7293\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.7467\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e0.7598\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1.00\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6878\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6793\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6575\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.6692\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.7214\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.8010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6672\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6520\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6417\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.6410\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.6431\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1.00\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7280\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.7017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.7684\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.7444\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.8492\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.7861\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7194\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.7434\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.7494\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.7204\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of different classifiers using Mobilenetv2\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClassifier\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTest accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1 measure\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9799\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9099\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9501\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9594\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9587\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.9603\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1.00\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8966\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9486\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9481\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9427\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.9603\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9671\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8949\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9456\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9489\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9579\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.9508\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e1.00\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9204\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.9658\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.9662\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.9660\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e0.9700\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.9850\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.9259\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9317\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9334\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9416\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.9246\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFeature extraction using Mobilenetv2 is performed better than InceptionV3, because InceptionV3 has more complex structure and has more hidden layers than MobileNetV2. As results shown, it performed comparatively better in training datasets but when it comes to validation and test accuracy its performance is not adequate. Based on the results shown in Table \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Table \u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, we can conclude based on results got from (BUSI) that when the datasets are limited then architectures having simple structure like MobileNetV2 come in handy and valuable for getting desired results. Among all the classifiers used in the experiments of this study ANN performed better than other classifiers. The reason behind the selection of these specific classifiers is that they are widely used in the literature of deep learning related problems [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e] [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e] [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. Moreover, this decision also helps to make the study robust in nature.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5.\tDISCUSSION ","content":"\u003cp\u003eThe present paper proposes a novel breast tumor classification scheme for ultrasound images based on combining feature extraction using InceptionV3 and MobileNetV2 with segmentation information. This study presents a comprehensive examination of breast cancer segmentation-based classification using advanced deep learning techniques. In contrast to previously published works related to deep learning-based breast classification works which only extract the features on breast ultrasound images, our work aims to combine medical knowledge in the network and design the segmentation network to obtain the tumor contour and region to extract more effective features for classification. The utilization of U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;architecture for segmentation and the combined application of InceptionV3 and MobileNetV2 for feature extraction have demonstrated promising results. The results show that our method has a better effect on breast tumor classification than other methods as shown in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Compared with other breast classification work Srikanthamurthy et al [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], including the transfer learning methods based on pre-trained models and shallow network methods, our method obtains better classification performance and the highest precision and accuracy, indicating that our method can classify breast tumor more effectively. The precision and recall metrics are shown in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, which also demonstrate a notable improvement in accuracy when utilizing deep learning-based classifiers, with the ANN classifier outperforming traditional methods like SVM and Random Forest. This supports the hypothesis that deep learning classifiers can capture complex Jabeen et al, Dar et al [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] as shown in Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, non-linear relationships within medical image data more effectively than their traditional counterparts. The performance of every aggregation method, including directly concatenating and directly adding feature representation of the two networks, is better than that of a single branch network, indicating that adding a parallel branch based on segmentation prior is meaningful for breast classification. This confirms that the feature aggregation network based on attention can effectively preserve the effective features and remove the redundant features. Xu et al [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e] proposed RMTL-Net for the simultaneously segmentation and classification of BUSI dataset and achieved accuracy of 91.18% in their model.\u003c/p\u003e \u003cp\u003eCompared with other feature aggregation methods, this method has a higher performance, demonstrating that our work can extract more effective and classification-related features when compared to Yaozhong et al [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e] study achieved recall value 91.18%. Arnab et al [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e] used mask images instead of segmented images for that reason the recall parameter outperformed the other state of the art for classification and achieved 96.95 %. Wheres, the segmentation is automatic in our scheme, which can enhance the operability of the system. Our work can provide an effective auxiliary diagnostic tool for the classification of BUSI and help effectively identify tumors in early screening. The study has enlightening implications for breast ultrasound classification and even other types of medical image diagnosis since more effective feature extraction is the basis for further image analysis.\u003c/p\u003e \u003cp\u003eWhile the results are promising, there are limitations that must be acknowledged. The dataset used, while sufficiently varied, is limited in size. Future work should aim to include a larger and more diverse set of images to ensure that the model can generalize across different populations and imaging conditions. Furthermore, the current study does not account for the real-time analysis of ultrasound images, a feature that would be necessary for practical clinical application. Additionally, while our results suggest that MobileNetV2 outperformed InceptionV3 in feature extraction, this may not hold across all datasets or tumor characteristics. The computational efficiency of MobileNetV2, however, makes it an attractive option for real-world applications where resources are limited. The implications of this research are significant, suggesting a path forward where AI can assist radiologists by providing a second opinion and reducing the workload associated with the segmentation and classification of breast tumors. The integration of such AI-based tools into clinical practice could potentially increase the early detection rates of breast cancer, thus improving patient outcomes.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of our model with state-of-the-art methods for BUSI datasets\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReferences\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMethodology\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 \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMeng Xu et al [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRMTL-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e93.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e93.34\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYaozhong et al [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSegmentation to Classification\u003c/p\u003e \u003cp\u003eScheme\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e90.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e91.18\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArnab et al [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClassification with DCNN(k\u0026thinsp;=\u0026thinsp;3) using binary masks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e96.95\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOur model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSegmentation-feature extraction-classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e96.58 (test)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e97\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"6.\tCONCLUSION","content":"\u003cp\u003eIn this research endeavor, we embarked on a journey to enhance the precision and effectiveness of breast cancer diagnosis through the integration of advanced deep learning techniques and classification models. The objective was to provide a robust and reliable framework that assists medical practitioners in early breast cancer detection, improving patient outcomes and treatment strategies.\u003c/p\u003e \u003cp\u003eOur comprehensive framework began with the segmentation of breast ultrasound images, where we harnessed the power of U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;to delineate regions of interest. The adoption of segmentation algorithms not only improved the accuracy of breast mass profiling but also paved the way for subsequent feature extraction.\u003c/p\u003e \u003cp\u003eFeature extraction, a pivotal step in our methodology, was carried out using both InceptionV3 and MobileNetV2 architectures. These feature-rich vectors encapsulated intricate details and characteristics within the segmented BUSI images. The fusion of features from these models enriched our feature space, enhancing our diagnostic capabilities.\u003c/p\u003e \u003cp\u003eThe subsequent classification phase was a testament to our commitment to thorough evaluation. Leveraging a diverse ensemble of classifiers, including Support Vector Machines, Artificial Neural Networks, k-Nearest Neighbours, Random Forest, and Long Short-Term Memory networks, we explored multiple avenues for accurate diagnosis. Each classifier brought its own strengths and nuances to the diagnostic task, allowing us to gain a comprehensive understanding of their capabilities.\u003c/p\u003e \u003cp\u003e Our extensive experiments, guided by rigorous evaluation metrics, provided valuable insights into the strengths and limitations of each model. Through this analysis, we aimed to empower medical practitioners with a suite of tools that can aid in their decision-making process. In conclusion, our research contributes to the field of breast cancer diagnosis by presenting a holistic framework that amalgamates the prowess of deep learning, image analysis, and classification techniques. While no single model emerged as the panacea, our ensemble of classifiers provides a versatile toolkit that can be tailored to specific diagnostic scenarios and requirements. As we reflect on our journey, we acknowledge that the path to improving breast cancer diagnosis is an ongoing one. Future endeavors may explore additional modalities, extend datasets, and fine-tune model architectures [\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e]. Our hope is that this research serves as a steppingstone, inspiring further innovation and collaboration in the pursuit of early breast cancer detection and improved patient care.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003edata availability\u003c/p\u003e\n\u003cp\u003eThe data supports the findings of this study is publicly available. Codes will be provided after acceptance and on request.\u003c/p\u003e\n\u003cp\u003eEthics declarations\u003c/p\u003e\n\u003cp\u003eConflict of interest\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003cp\u003eEthical standards\u003c/p\u003e\n\u003cp\u003eThis study did not use animal experiments.\u003c/p\u003e\n\u003cp\u003eData availability\u003c/p\u003e\n\u003cp\u003eThe data supports the findings of this study is publicly available. Codes will be provided after acceptance and on request.\u003c/p\u003e\n\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003e[Online].Available:https://gco.iarc.fr/tomorrow/en/dataviz/trends?cancers=20\u0026amp;scale=linear\u0026amp;min_zero=1.\u003c/li\u003e\n\u003cli\u003e[Online]. Available: https://www.cancer.org/research/acs-research-news/facts-and-figures-2022.html .\u003c/li\u003e\n\u003cli\u003eB. Malik and J. C. Klock, \u0026quot;Breast cyst fluid analysis correlations with speed of sound using transmission ultrasound,\u0026quot; \u003cem\u003eAcademic radiology, \u003c/em\u003evol. 26, no. 1, pp. 76-85, 2019. \u003c/li\u003e\n\u003cli\u003eM. Hu, Y. Li and X. Yang, \u0026quot;Breastsam: A study of segment anything model for breast tumor detection in ultrasound images,\u0026quot; \u003cem\u003earXiv preprint arXiv, \u003c/em\u003evol. 23, no. 05, 2023. \u003c/li\u003e\n\u003cli\u003eQ. Huang, F. Zhang and X. Li, \u0026quot;Machine Learning in Ultrasound Computer-Aided Diagnostic Systems:,\u0026quot; \u003cem\u003eBiomed Res Int., \u003c/em\u003e2018. \u003c/li\u003e\n\u003cli\u003eA. E.Ilesanmi, U. Chaumrattanakul and S. S. Makhanov, \u0026quot;Methods for the segmentation and classification of breast ultrasound images: a review.,\u0026quot; \u003cem\u003eJournal of Ultrasound, \u003c/em\u003evol. 24, pp. 367-382, 2021. \u003c/li\u003e\n\u003cli\u003eJ. Chowdary, P. Yogarajah, P. Chaurasia and V. Guruviah, \u0026quot;A Multi-Task Learning Framework for Automated Segmentation and Classification of Breast Tumors From Ultrasound Images,\u0026quot; \u003cem\u003eSage Journals, Ultrasonic Imaging, \u003c/em\u003evol. 44, no. 1, pp. 3-12, 2022. \u003c/li\u003e\n\u003cli\u003eP. Shi, J. Z. A. Rampun and H. Wang, \u0026quot;A hierarchical pipeline for breast boundary segmentation and calcification detection in mammograms.,\u0026quot; \u003cem\u003e Computers in biology and medicine, \u003c/em\u003evol. 1, no. 96, pp. 178-188, 2018. \u003c/li\u003e\n\u003cli\u003eR. Karmakar, S. Chatterjee, A. K. Das and A. Mandal, \u0026quot;BCPUML: Breast Cancer Prediction Using Machine Learning Approach\u0026mdash;A Performance Analysis.,\u0026quot; \u003cem\u003eSN Computer Science,, \u003c/em\u003evol. 4, no. 4, 2023. \u003c/li\u003e\n\u003cli\u003eS. Liu, Y. Wang, X. Yang, B. Lei, L. Liu, S. X. Li, D. Ni and T. Wang, \u0026quot;Deep learning in medical ultrasound analysis: a review,\u0026quot; \u003cem\u003eEngineering, \u003c/em\u003evol. 5, no. 2, pp. 261-275, 2019. \u003c/li\u003e\n\u003cli\u003eM. S. K. Inan, F. I. Alam and R. Hasan, \u0026quot;Deep integrated pipeline of segmentation guided classification of breast cancer from ultrasound images,\u0026quot; \u003cem\u003eBiomedical Signal Processing and Control, \u003c/em\u003evol. 75, 2022. \u003c/li\u003e\n\u003cli\u003eR. J. G. V. Sloun, R. Cohen and Y. C. Eldar, \u0026quot;Deep Learning in Ultrasound Imaging,\u0026quot; \u003cem\u003eProceedings of the IEEE, \u003c/em\u003evol. 108, no. 1, pp. 11-29, 2020. \u003c/li\u003e\n\u003cli\u003eY. Xu, Y. Wang, J. Yuan, Q. Cheng, X. Wang and P. L. Carson, \u0026quot;Medical breast ultrasound image segmentation by machine learning,\u0026quot; \u003cem\u003eUltrasonics, \u003c/em\u003evol. 91, pp. 1-9, 2019. \u003c/li\u003e\n\u003cli\u003eD. Perdios, M. Vonlanthen, F. Martinez, M. Arditi and J. P. Thiran, \u0026quot;CNN-based image reconstruction method for ultrafast ultrasound imaging. IEEE Transactions on Ultrasonics,\u0026quot; \u003cem\u003eIEEE Transactions on Ultrasound, Ferroelectrics and Frequency Control, \u003c/em\u003evol. 69, no. 4, pp. 1154-1168, 2022. \u003c/li\u003e\n\u003cli\u003eY. Peng, M. Sonka and D. Z. Chen, \u0026quot;U-Net v2: Rethinking the Skip Connections of U-Net for Medical Image Segmentation.,\u0026quot; \u003cem\u003earXiv preprint arXiv, \u003c/em\u003e2023. \u003c/li\u003e\n\u003cli\u003eN. Vimala and N. Anupama, \u0026quot;Precise segmentation of fetal head in ultrasound images using improved U‐Net model,\u0026quot; \u003cem\u003eETRI Journal, \u003c/em\u003e2023. \u003c/li\u003e\n\u003cli\u003eZ. Zhou, M. M. R. Siddiquee, N. Tajbakhsh and J. Liang, \u0026quot;UNet++: A Nested U-Net Architecture for Medical Image Segmentation,\u0026quot; in \u003cem\u003eSpringer International PublishinG\u003c/em\u003e, Granada, 2018. \u003c/li\u003e\n\u003cli\u003eG. Girish, P. Spandana and B. Vasu, \u0026quot;Breast cancer detection using deep learning,\u0026quot; \u003cem\u003eComputer Vision and Pattern Recognition, \u003c/em\u003e2023. \u003c/li\u003e\n\u003cli\u003eM. F. Dar and A. Ganivada, \u0026quot;EfficientU-net: a novel deep learning method for breast tumor segmentation and classification in ultrasound images,\u0026quot; \u003cem\u003eNeural Processing Letters, \u003c/em\u003evol. 55, pp. 10439-10462, 2023. \u003c/li\u003e\n\u003cli\u003eP. Pramanik, R. Pramanik, F. Schwenker and R. Sarkar, \u0026quot;DBU-Net: Dual branch U-Net for tumor segmentation in breast ultrasound images,\u0026quot; \u003cem\u003ePlos one, \u003c/em\u003evol. 18, no. 11, 2023. \u003c/li\u003e\n\u003cli\u003eN. S. Punn and S. Agarwal, \u0026quot;RCA-IUnet: a residual cross-spatial attention-guided inception U-Net model for tumor segmentation in breast ultrasound imaging,\u0026quot; \u003cem\u003eMachine Vision and Applications, \u003c/em\u003evol. 33, no. 2, 2022. \u003c/li\u003e\n\u003cli\u003eT. Zhao and H. Dai, \u0026quot;Breast tumor ultrasound image segmentation method based on improved residual u-net network.,\u0026quot; \u003cem\u003eComputational Intelligence and Neuroscience, \u003c/em\u003e2022. \u003c/li\u003e\n\u003cli\u003eK. Jabeen, M. A. Khan, M. Alhaisoni, U. Tariq, Y.-D. Zhang, A. Hamza, A. Mickus and R. Dama\u0026scaron;evičius, \u0026quot;Breast cancer classification from ultrasound images using probability-based optimal deep learning feature fusion,\u0026quot; \u003cem\u003eSensors, \u003c/em\u003evol. 22, no. 3, 2022. \u003c/li\u003e\n\u003cli\u003eZ. Zhuang, W. Ding, S. Zhuang, A. N. J. Raj, J. Wang, W. Zhou and C. Wei, \u0026quot;Tumor classification in automated breast ultrasound (ABUS) based on a modified extracting feature network. ,\u0026quot; \u003cem\u003eComputerized Medical Imaging and Graphics, \u003c/em\u003evol. 90, 2021. \u003c/li\u003e\n\u003cli\u003eM. M. Srikantamurthy, V. P. S. Rallabandi, D. B. Dudekula, S. Natarajan and J. Park, \u0026quot;Classification of benign and malignant subtypes of breast cancer histopathology imaging using hybrid CNN-LSTM based transfer learning.,\u0026quot; \u003cem\u003eBMC Medical Imaging, \u003c/em\u003evol. 23, no. 1, p. 19, 2023. \u003c/li\u003e\n\u003cli\u003eZ. Cao, L. Duan, G. Yang, T. Yue and Q. Chen, \u0026quot;An experimental study on breast lesion detection and classification from ultrasound images using deep learning architectures.,\u0026quot; \u003cem\u003eBMC medical imaging, \u003c/em\u003evol. 19, no. 1, pp. 1-9, 2019. \u003c/li\u003e\n\u003cli\u003eM. H. Yap, M. Goyal, F. M. Osman, R. Mart\u0026iacute;, E. Denton, A. Juette and R. Zwiggelaar, \u0026quot;Breast ultrasound lesions recognition: end-to-end deep learning approaches,\u0026quot; \u003cem\u003eJournal of medical imaging, \u003c/em\u003evol. 6, no. 1, 2018. \u003c/li\u003e\n\u003cli\u003eA. S. Podda, R. Balia, S. Barra, S. Carta, G. Fenu and L. Piano, \u0026quot;Fully-automated deep learning pipeline for segmentation and classification of breast ultrasound images,\u0026quot; \u003cem\u003eJournal of computational science, \u003c/em\u003evol. 63, 2022. \u003c/li\u003e\n\u003cli\u003eO. Ronneberger, P. Fischer and T. Brox, \u0026quot;U-Net: Convolutional Networks for Biomedical Image Segmentation,\u0026quot; \u003cem\u003eMedical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, \u003c/em\u003evol. 9351, pp. 234-241, 2015. \u003c/li\u003e\n\u003cli\u003eZ. Li, H. Zhang, Z. Li and Z. Ren, \u0026quot;Residual-Attention UNet++: A Nested Residual-Attention U-Net for Medical Image Segmentation,\u0026quot; \u003cem\u003eApplied Sciences, \u003c/em\u003evol. 12, no. 14, 2022. \u003c/li\u003e\n\u003cli\u003eP. Zhang, Z. Ma, Y. Zhang, X. Chen and G. Wang, \u0026quot;Improved inception V3 method and its effect on radiologists\u0026rsquo; performance of tumor classification with automated breast ultrasound system,\u0026quot; \u003cem\u003eGland Surgery, \u003c/em\u003evol. 10, no. 7, pp. 2232-2245, 2021. \u003c/li\u003e\n\u003cli\u003eT. Fujioka, M. Mori, K. Kubota, J. Oyama, E. Yamaga, Y. Yashima, L. Katsuta, K. Nomura, M. Nara, G. Oda, T. Nakagawa, Y. Kitazume and U. Tateishi, \u0026quot;The utility of deep learning in breast ultrasonic imaging: a review,\u0026quot; \u003cem\u003eDiagnostics, \u003c/em\u003evol. 10, no. 12, 2020. \u003c/li\u003e\n\u003cli\u003eN. Mohiuddin, R. A. Dar, M. Rasool and A. Assad, \u0026quot;Breast cancer detection using deep learning: Datasets, methods, and challenges ahead,\u0026quot; \u003cem\u003eComputers in biology and medicine, \u003c/em\u003evol. 149, 2022. \u003c/li\u003e\n\u003cli\u003eK. Dong, C. Zhou, Y. Ruan and Y. Li, \u0026quot;MobileNetV2 Model for Image Classification,\u0026quot; in \u003cem\u003e2020 2nd International Conference on Information Technology and Computer Application (ITCA)\u003c/em\u003e, Guangzhou, 2021. \u003c/li\u003e\n\u003cli\u003eM.-W. Huang, C.-W. Chen, W.-C. Lin, S.-W. Ke and C.-F. Tsai, \u0026quot;SVM and SVM ensembles in breast cancer prediction,\u0026quot; \u003cem\u003ePloS one, \u003c/em\u003evol. 12, no. 1, 2017. \u003c/li\u003e\n\u003cli\u003eS. Guizani, N. Guizani and S. Gharsallaoui, \u0026quot;A Hybrid CNN-SVM Prediction Approach for Breast Cancer Ultrasound Imaging,\u0026quot; \u003cem\u003eInternational Wireless Communications and Mobile Computing (IWCMC), \u003c/em\u003epp. 1174-1578, 2023. \u003c/li\u003e\n\u003cli\u003eA. A. Hossain, J. K. Nisha and F. Johora, \u0026quot;Breast cancer classification from ultrasound images using VGG16 model based transfer learning,\u0026quot; \u003cem\u003eInternational Journal of Image, Graphics and Signal Processing, \u003c/em\u003evol. 13, no. 1, 2023. \u003c/li\u003e\n\u003cli\u003eZ. Khandezamin, M. Naderan and M. J. Rashti, \u0026quot;Detection and classification of breast cancer using logistic regression feature selection and GMDH classifier,\u0026quot; \u003cem\u003eJournal of Biomedical Informatics, \u003c/em\u003evol. 111, 2020. \u003c/li\u003e\n\u003cli\u003eM. Bansal, A. Goyal and A. Choudhary, \u0026quot;A comparative analysis of K-Nearest Neighbor, Genetic, Support Vector Machine, Decision Tree, and Long Short Term Memory algorithms in machine learning,\u0026quot; \u003cem\u003eDecision Analytics Journal, \u003c/em\u003evol. 3, 2022. \u003c/li\u003e\n\u003cli\u003eH. Işik and S. Arslan, \u0026quot;An Artificial Neural Network classification approach for use the ultrasound in physiotherapy,\u0026quot; \u003cem\u003eNational Center for Biotechnology Information, \u003c/em\u003evol. 35, no. 6, 2011. \u003c/li\u003e\n\u003cli\u003eJ. G. Precious, S. Selvan and R. Avudaiammal, \u0026quot;Classification of abnormalities in breast ultrasound images using ANN, FIS and ANFIS classifier: A comparison,\u0026quot; in \u003cem\u003eIOP publishing\u003c/em\u003e, coimbatore, 2021. \u003c/li\u003e\n\u003cli\u003eC. V. Narasimhulu, \u0026quot;An automatic feature selection and classification framework for analyzing ultrasound kidney images using dragonfly algorithm and random forest classifier,\u0026quot; \u003cem\u003eIET Image Processing, \u003c/em\u003evol. 15, no. 9, pp. 2080-2096, 2021. \u003c/li\u003e\n\u003cli\u003eS. Zhang, M. Liao, J. Wang, Y. Zhu, Y. Zhang, J. Zhang, R. Zheng, L. Lv, D. Zhu, H. Chen and W. Wang, \u0026quot;Fully automatic tumor segmentation of breast ultrasound images with deep learning. ,\u0026quot; \u003cem\u003eJournal of Applied Clinical Medical Physics, \u003c/em\u003evol. 24, no. 1, 2023. \u003c/li\u003e\n\u003cli\u003eM. Radak, H. Y. Lafta and H. Fallahi, \u0026quot;Machine learning and deep learning techniques for breast cancer diagnosis and classification: a comprehensive review of medical imaging studies,\u0026quot; \u003cem\u003eJournal of Cancer Research and Clinical Oncology, \u003c/em\u003epp. 1-19, 2023. \u003c/li\u003e\n\u003cli\u003eW. Al-Dhabyani, M. Gomma, K. Hussein and A. Fahmy, \u0026ldquo;Dataset of breast ultrasound images,\u0026rdquo; \u003cem\u003eData in brief, \u003c/em\u003evol. 28, 2020. \u003c/li\u003e\n\u003cli\u003eM. Alruily, W. Said, A. M. Mostafa, M. Ezz and M. Elmezain, \u0026quot;Breast Ultrasound Images Augmentation and Segmentation Using GAN with Identity Block and Modified U-Net 3+,\u0026quot; \u003cem\u003eSensors, \u003c/em\u003evol. 23, no. 20, 2023. \u003c/li\u003e\n\u003cli\u003eM. Amiri, R. Brooks, B. Behboodi and H. Rivaz, \u0026quot;Two-stage ultrasound image segmentation using U-Net and test time augmentation,\u0026quot; \u003cem\u003eInternational journal of computer assisted radiology and surgery, \u003c/em\u003evol. 15, no. 6, pp. 981-988, 2020. \u003c/li\u003e\n\u003cli\u003eM. Mendez, S. Sundararaman, L. Probyn and P. N. Tyrrell, \u0026quot;Approaches and Limitations of Machine Learning for Synthetic Ultrasound Generation: A Scoping Review,\u0026quot; \u003cem\u003eJournal of Ultrasound in Medicine, \u003c/em\u003evol. 42, no. 12, pp. 2695-2706, 2023. \u003c/li\u003e\n\u003cli\u003eJ. Lian, M. Zhang, N. Jiang, W. Bi and X. Dong, \u0026quot;Feature extraction of kidney tissue image based on ultrasound image segmentation,\u0026quot; \u003cem\u003eJournal of Healthcare Engineering, \u003c/em\u003e2021. \u003c/li\u003e\n\u003cli\u003eMadhu and R. Kumar, \u0026quot;A hybrid feature extraction technique for content based medical image retrieval using segmentation and clustering techniques.,\u0026quot; \u003cem\u003eMultimedia Tools and Applications, \u003c/em\u003evol. 81, no. 6, pp. 8871-8904, 2022. \u003c/li\u003e\n\u003cli\u003eM. F. Mridha, M. A. Hamid, M. M. Monowar, A. J. Keya, A. Q. Ohi, M. R. Islam and J.-M. Kim, \u0026quot;A comprehensive survey on deep-learning-based breast cancer diagnosis,\u0026quot; \u003cem\u003eCancers, \u003c/em\u003evol. 13, no. 23, 2021. \u003c/li\u003e\n\u003cli\u003eM. Botlagunta, M. D. Botlagunta, M. B. Myneni, D. Lakshmi, A. Nayyar, J. S. Gullapalli and M. A. Shah, \u0026quot;Classification and diagnostic prediction of breast cancer metastasis on clinical data using machine learning algorithms,\u0026quot; \u003cem\u003eScientific Reports, \u003c/em\u003evol. 13, no. 1, 2023. \u003c/li\u003e\n\u003cli\u003eN. Saffari, H. A. Rashwan, M. Abdel-Nasser, V. K. Singh, M. Arenas, E. Mangina, B. Herrera and D. Puig, \u0026quot;Fully automated breast density segmentation and classification using deep learning.,\u0026quot; \u003cem\u003eDiagnostics, \u003c/em\u003evol. 10, no. 11, 2020. \u003c/li\u003e\n\u003cli\u003eD. R. Sarvamangala and R. V. Kulkarni, \u0026quot;Convolutional neural networks in medical image understanding: a survey,\u0026quot; \u003cem\u003eEvolutionary intelligence,, \u003c/em\u003evol. 15, no. 1, pp. 1-22, 2022. \u003c/li\u003e\n\u003cli\u003eM. Xu, K. Huang and X. Qi, \u0026quot;A Regional-Attentive Multi-Task Learning Framework for Breast Ultrasound Image Segmentation and Classification,\u0026quot; \u003cem\u003eIEEE Access, \u003c/em\u003evol. 11, pp. 5377-5392, 2023. \u003c/li\u003e\n\u003cli\u003eY. Luo, Q. Huang and X. Li, \u0026quot;Segmentation information with attention integration for classification of breast tumor in ultrasound image,\u0026quot; \u003cem\u003ePattern Recognition, \u003c/em\u003evol. 124, 2022. \u003c/li\u003e\n\u003cli\u003eArnab Kumar Mishra, S. B. Pinki Roy and S. K. Das, \u0026quot;Achieving highly efficient breast ultrasound tumor classification with deep convolutional neural networks,\u0026quot; \u003cem\u003eInternational Journal of information technology, \u003c/em\u003evol. 12, pp. 3311-3320, 2022. \u003c/li\u003e\n\u003cli\u003eH. Afrin, N. B. Larson, M. Fatemi and A. Alizad, \u0026quot;Deep Learning in Different Ultrasound Methods for Breast Cancer, from Diagnosis to Prognosis : Current Trends, Challenges, and an Analysis,\u0026quot; \u003cem\u003eCancers, \u003c/em\u003evol. 15, no. 12, pp. 31-39, 2023. \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","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":"Classification, MobilenetV2, InceptionV3, Segmentation, Feature extraction","lastPublishedDoi":"10.21203/rs.3.rs-3930759/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3930759/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eBreast cancer remains a significant global health challenge, demanding accurate and effective diagnostic methods for timely treatment. Ultrasound imaging stands out as a valuable diagnostic tool for breast cancer due to its affordability, accessibility, and non-ionizing radiation properties.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe evaluate the proposed method using a publicly available breast ultrasound images. This paper introduces a novel approach to classifying breast ultrasound images based on segmentation and feature extraction algorithm. The proposed methodology involves several key steps. Firstly, breast ultrasound images undergo preprocessing to enhance image quality and eliminate potential noise. Subsequently, a U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;is applied for the segmentation. A classification model is then trained and validated after extracting features by using Mobilenetv2 and Inceptionv3 of segmented images. This model utilizes modern machine learning and deep learning techniques to distinguish between malignant and benign breast masses. Classification performance is assessed using quantitative metrics, including recall, precision and accuracy. Our results demonstrate improved precision and consistency compared to classification approaches that do not incorporate segmentation and feature extraction. Feature extraction using InceptionV3 and MobileNetV2 showed high accuracy, with MobileNetV2 outperforming InceptionV3 across various classifiers.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe ANN classifier, when used with MobileNetV2, demonstrated a significant increase in test accuracy (0.9658) compared to InceptionV3 (0.7280). In summary, our findings suggest that the integration of segmentation techniques and feature extraction has the potential to enhance classification algorithms for breast cancer ultrasound images.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis approach holds promise for supporting radiologists, enhancing diagnostic accuracy, and ultimately improving outcomes for breast cancer patients. In future our focus will be to use comprehensive datasets to validate our methodology.\u003c/p\u003e","manuscriptTitle":"Deep Learning-Enhanced Ultrasound Analysis: Classifying Breast Tumors using Segmentation and Feature Extraction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-02-09 17:05:51","doi":"10.21203/rs.3.rs-3930759/v1","editorialEvents":[{"type":"communityComments","content":1}],"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":"e66c87be-87d8-4fbb-b89e-2c546f603836","owner":[],"postedDate":"February 9th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-04-15T16:58:25+00:00","versionOfRecord":[],"versionCreatedAt":"2024-02-09 17:05:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3930759","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3930759","identity":"rs-3930759","version":["v1"]},"buildId":"cTy_lsJlmDsVRNrSptgXS","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}