Explainable Deep Learning Framework for Accurate Breast Cancer Classification and Abnormality Detection from Mammogram Images

Abstract

Early and accurate detection of breast cancer is critical to improving outcomes for affected individuals. Although a mammogram is an important diagnostic tool, the examination requires a professional, and this will be time-consuming and inconsistent. Unbalanced datasets can make it hard for traditional DL methods to work, which lowers the accuracy and trustworthiness of classification. This paper is based on the KAU-BCMD Mammography dataset where mammogram images are labeled and the comments within the bounding box are marked for parts that are not normal. The entire dataset and training groups are provided with increased data as part of preprocessing, and the labels are encoded, features are extracted using ResNet50, and SMOTE is employed for class balancing.

Problems are identified using YOLOv5, YOLOv8 and YOLOv11. The accepted classifiers are ResNet50 with FC-MLP, Xception, and Hybrid Xception-ResNet50. The range was completed using Accuracy, Precision, Recall, F1 score, ROC AUC and modes have been tested. The metrics used for detection were mAP, Precision, and Recall. The average of all the metrics indicates that the proposed_full_bal classification model achieved 99.5% and YOLOv8 had the highest detection success with mAP 0.462. Heatmaps were created using Explainable AI techniques like Grad-CAM to elucidate the regions of the image where the model decision-making process plays a crucial role. A Flask-based interface is provided that allows the user to upload mammogram images and receive predicted classes and boxes. These results indicate that the use of a mixture of deep feature extraction, balanced registration, hybrid architecture, and interpretive visualization can increase the accuracy of most breast cancer diagnoses and classify mammogram use.