Pipeline for Visual Interpretability in Breast Cancer Classification Using Grad-CAM Zahra Azizah, Syahril Siregar
Politeknik Negeri Jakarta, Dept. of Computer and Informatics Engineering
Universitas Indonesia, Dept. of Physics, Faculty of Mathematics and Natural Sciences
Abstract
Breast cancer is a major health challenge, where early detection significantly improves outcomes. While convolutional neural networks (CNNs) achieve high accuracy in breast lesion classification, their lack of interpretability limits clinical adoption. This study presents a computer-aided pipeline for mammogram classification using the pre-trained ResNet-50 model for feature extraction and classification. Preprocessing steps include CLAHE for contrast enhancement, YOLO for Region of Interest detection, and data augmentation to mitigate limited dataset challenges. To address interpretability, Grad-CAM is integrated, providing feature-based visualizations aligned with expert-identified abnormalities. Our pipeline demonstrates robust performance on CC view images with high precision and accuracy while highlighting challenges in MLO view analysis due to variability and model sensitivity. Grad-CAM enhances transparency, supporting radiologists in validating predictions. This approach improves diagnostic accuracy and trustworthiness, paving the way for more effective breast cancer screening. Future efforts will focus on refining MLO view performance and validating across larger datasets.
Keywords: breast cancer, mammography, Convolutional Neural Networks (CNNs), pre-trained ResNet-50, Grad CAM, data preprocessing, medical image analysis
