A Deep Learning-Based Multi-Class Breast Cancer Histopathological Image Classification System Using CBAM

Abstract

Breast histopathological image classification is challenging because benign and malignant breast lesions can exhibit overlapping cellular morphology, texture, staining appearance, and tissue architecture. This paper presents a multi-class deep learning framework that integrates the BreakHis dataset, ImageNet-pretrained InceptionResNetV2, and the Convolutional Block Attention Module (CBAM) for eight-class histopathological image classification. The pipeline performs class-distribution analysis, mean-count-based oversampling and undersampling, CLAHE-based contrast enhancement in LAB space, resizing to 224×224 pixels, normalization, and horizontal and vertical training augmentation. InceptionResNetV2 provides the main convolutional representation, while CBAM sequentially applies channel and spatial attention to refine discriminative feature information before a dense classification head. The latest ten-epoch training run reaches 98.44% training accuracy. The reported test evaluation achieves 98% accuracy on 2,160 test images, with 270 images per class. Macro precision, recall, and F1-score are approximately 0.98. The results demonstrate strong performance for the integrated pipeline under the reported benchmark protocol. At the same time, the paper identifies methodological limitations concerning resampling before partitioning, image-wise rather than patient-wise splitting, the absence of component-level ablation, and the current hard-label ROC/AUC implementation. These limitations define the scope of the reported findings and motivate stricter patient-wise, external, and probability-based validation.