Favoring flat minima improves generalization of transfer-learned models for breast ultrasound tumor classification

Abstract

Transfer learning with convolutional neural networks (CNNs) for breast ultrasound tumor classification can be susceptible to overfitting and unstable generalization when training data are limited, noisy, and heterogeneous. Sharpness-Aware Minimization (SAM) is an optimizer-level training strategy that encourages convergence to flatter regions of the loss landscape and may improve held-out performance without modifying model architecture. In this study, we systematically evaluated SAM for breast ultrasound image classification across seventeen CNN architectures spanning five families-VGG/AlexNet, ResNet, DenseNet, MobileNet, and EfficientNet-using four experiments on two datasets (BUSI and BUS-UCLM). We compared standard Adam versus SAM+Adam and standard stochastic gradient descent (SGD) versus SAM+SGD under matched experimental settings. On BUSI, SAM improved mean validation accuracy by +2.30 percentage points with Adam and by +2.51 percentage points with SGD. On BUS-UCLM, the corresponding mean gains were +2.27 and +2.55 percentage points. The largest individual improvement was observed for MobileNetV3 Small under SAM+SGD (+4.86 percentage points on BUSI), while the highest validation accuracy was achieved by VGG19 with SAM+SGD (96.20% on BUSI). Improvements were also observed in precision, recall, and F1-score, and paired statistical testing across architectures showed consistently positive effects with large effect sizes. Computational analysis showed moderate training-time overhead but no architectural inference overhead, supporting the practical feasibility of SAM. Overall, these findings indicate that SAM is a useful and broadly effective training enhancement for transfer-learned breast ultrasound classification across multiple CNN families and two independent datasets.