In this research study, we introduce a novel system of two primary modules: (1) Computer-Aided Detection (CADe) and (2) Computer-Aided Diagnosis (CADx). The CADe module is dedicated to the detection and segmentation of potentially anomalous regions within mammograms through the utilization of a pre-trained YOLOv5-seg model. This approach facilitates the early detection of breast cancer. Subsequently, the CADx module leverages the identified and segmented regions for further analysis, employing the VGG 16 classification model to discern the benign or malignant nature of these regions. To gauge the eficacy of our proposed methodology, extensive experiments were conducted on a substantial dataset procured from the Digital Database for Screening Mammography (DDSM). The CADe system yielded robust results in terms of detection and segmentation assessments, with a mean average precision (mAP) of 88%, a precision rate of 93.93%, and a recall rate of 98.02%. Furthermore, the CADx system demonstrated an accuracy rate of 97%.
professionals employ various diagnostic tools,
including mammography, magnetic resonance imaging (MRI),
Cancer, characterized by uncontrolled cell growth and and biopsy, to diagnose and evaluate breast cancer [
When it comes to existing diagnostic methods, the radiologists identify breast cancer indicators on
mammoearly detection of breast cancer plays a crucial role in grams, and CADx, which aids in distinguishing between
improving patient outcomes. The origins of breast cancer diferent tumor types, encompassing the broader field of
predominantly result from genetic mutations, accounting computer-aided diagnostics.
for only around 10% of cases, while the majority occur Machine learning algorithms have made significant
spontaneously [
The structure of this paper is organized as follows. Sec- cies of 88.74%, 92.56%, and 95.32%, respectively.
tion 2 will delve into existing research and prior work A. Bal et al. [
Mefty Zahira Hanan et al. introduced CAD system 2. Related Work for breast cancer, employing CNN models [9]. Their approach encompasses two pivotal stages: detection (CADe) In recent years, significant advancements have been and identification (CADx). By meticulously fine-tuning made in the field of breast cancer detection and diagnosis, the VGG19 CNN model on the DDSM dataset for CADe thanks to notable progress in deep learning techniques. and the IDC dataset for CADx, they accomplished reMany studies have explored innovative approaches and markable results, attaining a striking accuracy of 99% for frameworks, all aimed at improving the accuracy and CADe and 91% for CADx. eficiency of (CAD) systems for breast lesion analysis. G. H. Aly et al. [10], introduce a breast cancer detection This section provides a comprehensive overview of the framework utilizing YOLO-V3 and YOLO-V4 models. The most notable contributions in this area. framework identifies masses in mammograms and cate
Yousefi kamal P. [
M. A. Al-antari et al. [
J.Shi. [13], utilizes a YOLO-based computer-aided di- YOLOv5-seg is a specialized variant of the YOLO (You agnosis (CAD) system to address the challenges associ- Only Look Once) [15], algorithm designed for precise inated with chest cancer detection. Three key issues are stance segmentation tasks. It directly predicts pixel-level discussed and analyzed within the CAD system imple- masks for objects in images by combining YOLOv5’s obmentation: the utilization of handcrafted features, the ject detection with ProtoNet, a fully connected network prevalent high false positive rate in clinical settings, and utilizing 2D convolutions with SiLU activation functions the complexity of detecting irregular nodules in spiral to generate mask prototypes. YOLOv5-seg surpasses CT scans. YOLOv5 in channel outputs, with 351 channels, due to
These related works demonstrate the growing interest its additional 32 mask outputs. This integration efecin utilizing deep learning methods, particularly CNNs, for tively enables instance segmentation by leveraging Probreast cancer detection and classification. They provide toNet instance features and object detection information, valuable insights into the advancements and potential making it suitable for real-time applications, including solutions in this field, contributing to the development medical imaging. Various weight variants (YOLOv5sof accurate and eficient CAD systems for breast cancer seg, YOLOv5m-seg, YOLOv5l-seg, and YOLOv5x-seg) diagnosis. are available to cater to diferent complexity and accuracy requirements [16].A depiction outlining the architectural arrangement of YOLOv5-sag is provided in Figure 3. Models 2.[15, 16].
3.1. VGG16 VGG16, a widely recognized deep CNN architecture introduced by Simonyan and Zisserman. [14], is celebrated Figure 2: Elucidation of the Comprehensive Structure of for its depth and eficacy in image classification tasks. It YOLOv5l-seg comprises 16 convolutional layers organized into blocks, each housing multiple 3x3-sized filter layers. This choice of filter size enables the model to discern intricate image features efectively. VGG16 maintains uniformity 4. Datasets by consistently using 3x3 filters, and it simplifies complexity with 2x2 pooling layers, reducing activation map In our work, we primarily utilized the Digital Database sizes while preserving vital features. After convolution for Screening Mammography (DDSM) [17], a widely recand pooling, the model flattens the features for classifica- ognized public database extensively employed in breast tion using softmax, making it adept at object recognition. cancer diagnostic and preventive assistance systems. The VGG16’s extensive depth facilitates the capture of com- DDSM dataset includes 2620 cases, each containing four plex patterns, enhancing its classification performance LJPEG-format images representing two images for each across various computer vision applications, including breast (Left_CC, Left_MLO, Right_CC, Right_MLO) from medical image analysis.A visual representation elucidat- diferent angles. In total, it comprises 10,480 images deing the architectural configuration of VGG16 is appended picting various breast conditions, including normal, cancerous, and benign states. For this study, 800 images were randomly selected from the DDSM dataset, with an equal distribution of 400 normal and 400 abnormal (cancerous and benign) cases.
our CNN model, a normalization process is employed. In this process, the entire images are resized to a standardized dimension of 200×200 pixels [17, 18]. This resizing step ensures uniformity in the input image size, allowing for consistent processing and analysis by the CNN model.
The choice of 200×200 pixels as the target dimension is based on considerations of both computational feasibility and preserving important image information. This dimension strikes a balance between capturing significant details within the breast tissue and maintaining a manageable computational workload.
Data splitting in deep learning involves partitioning the dataset into three subsets: training (70%), validation (10%), and test (20%). The training set is used for model training, the test set assesses performance, and the validation set ifne-tunes hyperparameters and guides model optimization.
and segmentation of medical images. This method incorporates state-of-the-art techniques, including CNN YOLOv5-seg for precise region of interest segmentation, and VGG16 for the final classification of results. This approach is designed to enhance the accuracy and eficiency of the Computer-Aided Diagnosis (CAD) system in the medical domain. In the following sections, we will delve into each component of our method and explain how they are integrated to achieve an overall high-performing solution.Figure 3. illustrates a comprehensive depiction of our methodology.
the YOLOv5L-seg variant, which exhibited the best performance for our specific database, and we made some changes to it. 6.1.1. Addition of convolutional layers
pixels), we introduced three additional convolutional layers into the detection head of YOLOv5-seg. This strategic augmentation was undertaken to extract finer-grained information. The specifications of these added layers are depicted in Table 1.
Table 2 facilitate the transfer of information extracted by the Evolution of Convolutional Layer Characteristics through Fil- convolutional layers to the fully connected layers. ters Modifications. Finally, the development of a new classifier represented the ultimate step. Situated above the existing convoluLayers CONV’1 CONV’2 CONV’3 tional layers, this classifier was composed of three speBefore 125 256 512 cific layers. The first layer was a fully connected layer After 256 512 1024 endowed with 256 neurons and activated by a ReLU function. The second layer was also fully connected and comprised 128 neurons activated by ReLU. The final output
These adaptations aim to broaden the receptive field of oriented layer featured a solitary neuron, its activation the filters and enhance the detection head’s capability to governed by a sigmoid function for binary classification. capture salient features across diferent spatial scales.The Consequently, this classifier yields output probabilicomplete modifications are depicted in Figure 4. ties through the sigmoid activation function. A value proximate to zero indicates a diminished probability of malignancy, while a value approaching one signifies a substantial probability of malignancy.The new architecture is presented in Figure 5.
7.1. Assessment Criteria For the YOLO model, we evaluate its performance by using mAP, precision, and recall. These metrics assess the In this section, we discuss the evaluation metrics used to model’s ability to accurately detect and classify objects assess the performance of our proposed method, specif- in the given dataset. ically for the YOLO model and the VGG16 model. We For the VGG16 model, we focus on accuracy as the focus on key metrics such as mean Average Precision primary evaluation metric. Accuracy provides a compre(mAP), precision, recall, and accuracy to provide a com- hensive measure of the model’s correctness in classifying prehensive evaluation of the models’ performance. images into their respective categories. • True Positives (TP) : It represents the number of positive instances correctly identified or 7.2. Outcome Summary classified as positive by a model or test. In this section, we present the outcomes of our research • True Negatives (TN) : It represents the num- study, which focuses on the combined use of CADe and ber of negative instances correctly identified or CADx techniques. Our experimental setup involved uticlassified as negative by a model or test. lizing a 10th generation Intel Core i7 processor coupled • False Positives (FP) : cIt represents the with an NVIDIA GeForce MX150 graphics card for trainnumber of negative instances incorrectly identi- ing and evaluation purposes. ifed or classified as positive by a model or test. In other words, it is the number of instances that are actually negative but were mistakenly classified 7.2.1. CADe Results as positive. For CADe, we employed the YOLOv5-seg model, con• False Negatives (FN) : It represents the num- ducting extensive experiments to achieve improved perber of positive instances incorrectly identified or formances. Over 100 iterations were conducted using classified as negative by a model or test. In other various YOLOv5-seg models, each with diferent hyperwords, it is the number of instances that are ac- parameters, aiming to enhance the results. The obtained tually positive but were mistakenly classified as results are presented in Figure 6. and Table 3. negative. • Accuracy : It is a common evaluation metric that measures the overall correctness of the predictions made by the model. • • • =
+ + + + Precision : Precision measures the accuracy of positive predictions made by the models.
=
+
Recall : Recall, also known as sensitivity or true positive rate, measures the model’s ability to correctly detect positive instances.
=
+
Mean Average Precision (mAP) : The computer vision research community relies on the Mean Average Precision (mAP) as a standard metric to assess the reliability of object detection models. It quantifies the performance of these models by calculating the mean of average precision (AP) values across recall values ranging from 0 to 1.
(1) (2) (3) () d mAP =
1 ∑︁ AP (4) =1 Table 3 primary phases: Detection (CADe) and Identification Performance Evaluation of YOLOv5l-seg Through Multiple (CADx). Recent years have seen a rapid ascent in the Training Iterations: A Comparative Analysis. prominence of deep learning as a robust method for predictive analysis, particularly in the realm of image proEPOCHS MODEL mAP cessing. Deep learning Convolutional Neural Network (CNN) models, renowned for their hierarchical architec40 YOLOv5m-seg1 22% tures, have proven invaluable in extracting intricate fea40 YOLOv5m-seg2 26% tures from input images. In the development of our CAD system, a meticulous selection process was employed to 100 YOLOv5l-seg1 49% identify pre-trained CNN models that exhibited promis100 YOLOv5l-seg2 68% ing outcomes. Specifically, YOLOv5-seg was designated 300 YOLOv5l-seg1 70% for CADe, while VGG16 emerged as the most suitable choice for CADx. 300 YOLOv5l-seg2 88% The principal objective of this research endeavor was to develop a Computer-Aided Diagnosis (CAD) system aimed at augmenting radiologists’ diagnostic capabili7.2.2. CADx Results ties, thereby ensuring heightened accuracy in patient assessments. This CAD system was structured into two Moving on to CADx, we employed the VGG116 model primary phases: Detection (CADe) and Identification and performed over 100 iterations for evaluation. After (CADx). Recent years have seen a rapid ascent in the 100 epochs of training, we achieved an impressive preci- prominence of deep learning as a robust method for presion of 97%.The obtained results are presented in Figure dictive analysis, particularly in the realm of image pro7. cessing. Deep learning Convolutional Neural Network (CNN) models, renowned for their hierarchical architectures, have proven invaluable in extracting intricate features from input images. In the development of our CAD system, a meticulous selection process was employed to identify pre-trained CNN models that exhibited promising outcomes. Specifically, YOLOv5-seg was designated for CADe, while VGG16 emerged as the most suitable choice for CADx.
While the strides made in developing an advanced Computer-Aided Diagnosis (CAD) system employing deep learning models such as YOLOv5-seg and VGG16 are noteworthy, it is imperative to acknowledge broader implications beyond performance metrics. One significant aspect deserving attention in deploying such systems at scale is the safeguarding of patient data security. Ensuring the integrity and confidentiality of medical information within the CAD system against potential breaches is crucial, aligning with stringent data protection regulaFigure 7: Displaying Training and Validation Accuracies for tions to uphold patient privacy.
VGG16 Furthermore, the real-world application of these systems necessitates consideration of latency issues, especially in time-sensitive scenarios like medical diagnoses.
Minimizing latency between image input and diagnosis 8. CONCLUSION output is pivotal for improving clinical workflow eficiency and providing timely insights to healthcare proThe principal objective of this research endeavor was fessionals. Achieving low latency in real-time processing to develop a Computer-Aided Diagnosis (CAD) system without compromising accuracy becomes essential to aimed at augmenting radiologists’ diagnostic capabili- ensure the practical usability of these systems. ties, thereby ensuring heightened accuracy in patient Another critical aspect revolves around the manageassessments. This CAD system was structured into two ment of extensive datasets in real-world scenarios. As 1Original YOLOv5-seg the CAD system operates within clinical environments, 2Modified YOLOv5-seg seamless handling and storage of vast datasets become imperative. Scalability and eficient data management India, January 7–10, 2021, Proceedings 17, Springer, techniques are vital to accommodate the continuous in- 2021, pp. 253–267. lfux of medical imaging data while preserving system [9] M. Z. Hanane, M. Mejdeded, Utilization of preperformance and responsiveness. trained models of cnn in mammograms processing
Addressing these challenges—data security, latency op- for the diagnosis of breast cancer, in: 2022 7th timization, and efective management of extensive real- International Conference on Image and Signal Proworld datasets—becomes integral for the successful in- cessing and their Applications (ISPA), IEEE, 2022, tegration and sustainable utilization of CAD systems in pp. 1–5. clinical settings. While our research demonstrated im- [10] G. H. Aly, M. A. E.-R. Marey, S. El-Sayed Amin, pressive accuracy and precision, future advancements M. F. Tolba, Yolo v3 and yolo v4 for masses detecshould focus not only on model performance but also on tion in mammograms with resnet and inception for overcoming these practical hurdles to ensure the seam- masses classification, in: Advanced Machine Learnless and secure deployment of CAD systems in enhancing ing Technologies and Applications: Proceedings of patient care and diagnosis. AMLTA 2021, Springer, 2021, pp. 145–153. [11] F. A. Zeiser, C. A. da Costa, T. Zonta, N. M. Marques, A. V. Roehe, M. Moreno, R. da Rosa Righi, References Segmentation of masses on mammograms using data augmentation and deep learning, Journal of Algeria, digital imaging 33 (2020) 858–868.
[12] H. Chougrad, H. Zouaki, O. Alheyane, Deep convolutional neural networks for breast cancer screening, Computer methods and programs in biomedicine 157 (2018) 19–30. [13] J. Shi, A technical comparison of yolo-based chest cancer diagnosis methods, Highlights in Science,
Engineering and Technology 41 (2023) 35–42. [14] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, arXiv preprint arXiv:1409.1556 (2014). [15] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779–788. [16] G. Jocher, A. Chaurasia, A. Stoken, J. Borovec,
Y. Kwon, K. Michael, J. Fang, Z. Yifu, C. Wong, D. Montes, et al., ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation, Zenodo (2022). [17] G. V. Suganthi, J. Sutha, M. Parvathy,
N. Muthamil Selvi, Genetic algorithm for feature selection in mammograms for breast masses classification, Computer Methods in Biomechanics and Biomedical Engineering: Imaging &
Visualization (2023) 1–12. [18] M. Heath, K. Bowyer, D. Kopans, P. Kegelmeyer Jr,
R. Moore, K. Chang, S. Munishkumaran, Current status of the digital database for screening mammography, in: Digital Mammography: Nijmegen, 1998, Springer, 1998, pp. 457–460.