Skip to content
Conference Open access

MobileNetV4-CBAM for Interpretable Multi-Class Breast Cancer Ultrasound Classification

Sep 2026 · 2026 International Conference on Data Science and Its Applications (ICoDSA) · pp. 172-177 · 0 citations · 30 references

Abstract

Breast cancer's high case burden and mortality among women make accurate diagnosis essential for clinical management. While breast ultrasound (US) provides real-time lesion characteristics, its interpretation remains affected by speckle noise, visual similarities among lesions, and examiner subjectivity. Deep learning can aid classification, but conventional models often fail to emphasize the most informative features and remain black-box in nature. Moreover, evaluations of MobileNetV4-CBAM on merged BUSI and Mendeley datasets for three-class classification, supported by multiple explainability methods, remain limited. To address this gap, this study presents MobileNetV4-CBAM, an approach that integrates the Convolutional Block Attention Module into MobileNetV4 to categorize 1,028 US scans as benign, malignant, or normal, with 103 images (10%) reserved as an independent test set. Following median filtering and CLAHE preprocessing, the proposed architecture was benchmarked against baseline models, including ViT, InceptionV3, and VGG variants, using a stratified fivefold cross-validation scheme and independent testing. Inferential transparency was further validated through Grad-CAM, LIME, and SHAP. The proposed architecture outperformed all baselines, achieving 93.20% accuracy, 95.12% precision, 92.50% recall, and a 93.71% F1-Score. Visual and feature-level analyses indicated more concentrated lesion-focused activation and discriminative feature utilization across classes. These results demonstrate that embedding attention mechanisms within lightweight architectures enhances both predictive performance and interpretability. Ultimately, this integration strengthens diagnostic accuracy while demystifying computational reasoning, contributing toward trustworthy and interpretable breast cancer screening tools.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.