Preliminary exploration on using entropy-weighted hybrid pooling in CNN for ultrasound breast cancer detection.

Introduction: Breast ultrasound imaging is valuable for its non-invasiveness and cost-effectiveness but presents diagnostic challenges from speckle noise and limited contrast. Conventional CNN pooling methods compromise between noise reduction and feature preservation. Methods: We introduce an adaptive entropy-weighted hybrid pooling approach that dynamically combines Max and Average pooling based on local image complexity measured by Shannon entropy. Two CNN architectures (3-block and 4-block) were evaluated on a publicly available ultrasound dataset of 9,016 images with an additional 10% speckle-noise variant, using accuracy, precision, recall, F1-score, ROC curves, confusion matrices, and inference time. For the 3-block CNN, results are reported as the mean ± standard deviation over three random seeds. Results: In the 3-block CNN, hybrid pooling achieved an accuracy of 93.98% ± 1.72% (AUC = 0.9870), exceeding max pooling (92.72% ± 0.85%, AUC = 0.9815). In the deeper 4-block CNN (single-run), hybrid pooling remained competitive (accuracy 92.90%) relative to max pooling (94.79%). Deeper architectures improved accuracy, noise robustness, and convergence but required careful regularization to avoid overfitting. Discussion: This preliminary study highlights the adaptive hybrid pooling's potential in clinical ultrasound breast cancer diagnosis, recommending further validation and clinical integration.