Comparação de Wavelet Pooling e Dual-Branch com CBAM e ROI para Classificação de Mamografias

Mammography is the primary breast cancer screening method, yet manual interpretation yields relevant false-negative rates. CNNs have shown promise for computer-aided diagnosis, but the isolated contribution of ROI extraction, attention, and wavelet pooling remains underexplored. This paper compares ten architectures on the mini-MIAS dataset via Stratified 5-Fold Cross-Validation, evaluating ROI extraction, CBAM, four wavelet families as pooling replacements, a hybrid strategy, and a dual-branch model (ResNet18 + Wavelet Coiflet). ROI extraction is the most critical component (+35.6 p.p.); Daubechies-4 is the most stable wavelet for pooling (Acc 92.81%, σ=0.77); and the dual-branch achieves the best performance (Acc 94.06%, F1 0.939, AUC 0.936, sensitivity 99.5%).