Enhancing breast lesion segmentation in ultrasound images using U-Net with augmentation-based preprocessing

Abstract

Accurate segmentation of breast lesions in ultrasound remains challenging due to limited data and acquisition variability. To quantify how common augmentation operators and their combinations affect cross-dataset generalization of a standard U-Net for breast lesion segmentation. We trained on the BrEaST dataset ( n = 256) and evaluated on held-out validation and an independent external test set (QAMEBI, n = 232). We ablated rotation, flips, scaling, translation, shear, elastic deformation, grid distortion, and brightness/contrast (Albumentations), and then assessed combined policies. Performance was measured by Jaccard Similarity Index (JSI) and recall. Multiple augmentations substantially improved performance over a basic preprocessing pipeline. On external test data, the best combined policy improved JSI from 73.94 to 79.36% and recall from 86.50 to 89.24 % vs. baseline; grid distortion, rotation, and vertical flips were most impactful, while shear showed inconsistent generalization. Carefully chosen augmentation policies provide a simple, reproducible way to boost out-of-distribution segmentation performance in breast ultrasound. We release augmentation recipes to support adoption.