Improving Anomaly Detection Techniques in Brest cancer ultrasound Using Convolutional Neural Networks

Breast ultrasound (BUS) is a widely used, non-ionizing imaging modality, partic- ularly valuable in dense breast tissue and resource-limited settings. However, auto- mated analysis remains challenging due to speckle noise, heterogeneous textures, and pronounced class imbalance between benign and malignant lesions. We propose a compact and reproducible deep learning pipeline based on EfficientNet-B0, enhanced with focal loss and class-balanced augmentations, for robust three-way classification ( normal / benign / malignant ). On the BUSI dataset, our method achieves a state- of-the-art 99.4% accuracy and a weighted F1-score of 0.99 . Beyond standard metrics, we perform a rigorous evaluation including bootstrap confidence intervals, paired t -tests, and McNemar ’s test. We assess reliability using Expected Calibration Error (ECE) and provide Grad-CAM visualizations to confirm that the model at- tends to clinically meaningful regions. Ablation studies quantify the impact of focal loss, fine-tuning depth, and optional CLAHE preprocessing. The proposed approach is efficient, interpretable, and deployment-ready, establishing a strong baseline for future BUS systems incorporating attention mechanisms or vision transformers.