Deep Learning-Based Segmentation, Classification, and Explainable Diagnosis of Breast Cancer Lesions

Abstract

Breast cancer remains one of the most prevalent malignancies affecting women worldwide, and early, accurate detection through medical imaging is central to improving survival outcomes. This paper presents an end-to-end computer-aided diagnosis (CAD) framework that performs lesion segmentation, region-of-interest (ROI) extraction, tumor stage estimation, benign/malignant classification, and visual explainability across two complementary imaging modalities: digital mammography, represented by the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM), and breast ultrasound, represented by the Breast Ultrasound Images (BUSI) dataset. For each modality, a modality-specific U-Net architecture was trained to generate binary lesion masks, from which tumor coverage percentage and a rule-based disease stage were derived. Three classifier backbones were benchmarked per modality — ResNet50, VGG16, and a custom convolutional neural network (CNN) for CBIS-DDSM (benign vs. malignant), and ResNet50, EfficientNet-B0, and a custom CNN for BUSI (benign vs. malignant vs. normal) — with Gradient-weighted Class Activation Mapping (Grad-CAM) applied to the strongest classifiers for interpretability. The trained pipelines were deployed in a Django-based web application allowing end users to upload an image and receive segmentation, ROI, stage, classification, confidence, and Grad-CAM outputs in real time. Results show a pronounced modality gap: the U-Net segmentation network achieved a validation Dice coefficient of 0.462 on BUSI versus 0.247 on CBIS-DDSM, and the best classifier reached 84.07% test accuracy on BUSI (ResNet50) compared with 61.86% on CBIS-DDSM (ResNet50), with markedly lower malignant-class recall on mammography (37%) than on ultrasound (76%). These findings quantify how imaging modality, lesion conspicuity, and dataset scale materially affect both segmentation and diagnostic performance, and they underscore the necessity of transparent, modality-aware reporting — including recall/sensitivity for the malignant class, not just aggregate accuracy — when evaluating deep learning systems intended to support breast cancer screening.