Adaptive Variance Pooling CNN for Noise-Robust Ultrasound Breast Cancer Classification

This study proposes a novel Weighted Variance-based Adaptive Average Pooling (WVAAP) method designed to enhance the robustness and diagnostic accuracy of convolutional neural networks (CNNs) in classifying breast cancer from ultrasound images. Unlike conventional pooling methods, WVAAP dynamically adjusts activation weights based on local variance, effectively balancing noise reduction and critical feature preservation. Evaluated on a publicly available dataset, the proposed method achieved superior accuracy (85.37%), precision (86.29%), recall (85.37%), and F1-score (85.27%) compared to traditional pooling methods. While WVAAP requires a modest inference time of 9.58 ms, this remains clinically acceptable for rapid diagnostic decision-making by healthcare professionals. Additionally, ROC curve analysis demonstrated strong discriminative performance (AUC = 0.9191), highlighting WVAAP’s potential for reliable clinical ultrasound breast cancer diagnosis.