Chaotic Learning Rate Scheduling for Improved CNN-Based Breast Cancer Ultrasound Classification
İshak Paçal
Abstract
Early and accurate identification of breast cancer from ultrasound is a key clinical objective, yet performance remains sensitive to how deep models are trained. Convolutional neural networks (CNNs) achieve strong results, but conventional learning rate schedulers follow rigid, deterministic trajectories that can limit exploration of complex loss landscapes. In this study, we introduce a Chaotic Learning Rate Scheduler (CLRS) that modulates the learning rate via the bounded, nonperiodic dynamics of the logistic map to encourage broader yet stable exploration during training. We assess CLRS in a controlled comparison on the BUSI dataset using four pretrained backbones. All components of the pipeline other than the scheduling policy are fixed, including data splits, preprocessing, optimizer settings, and model selection. Under an identical 300-epoch budget, CLRS is evaluated against a cosine scheduler, and test performance is reported using accuracy, precision, recall, and macro F1. LayerCAM is used to examine whether numerical gains correspond to clinically meaningful spatial attention. Across all backbones, CLRS consistently outperforms the cosine baseline. The best configuration, based on EfficientNetV2 Small, attains 0.9391 accuracy and 0.9255 macro F1. Gains are most pronounced in recall and macro F1, indicating improved sensitivity without additional inference cost. Parameter counts and GMACs remain unchanged, showing that benefits arise from altered training dynamics rather than model capacity. LayerCAM visualizations reveal more lesion-centric attention in correctly classified cases and support CLRS as an effective, deployment-neutral strategy for breast ultrasound classification.