Uncertainty-Calibrated Test-Time Adaptation for Cross-Dataset Breast Ultrasound Segmentation

Abstract

Breast ultrasound lesion segmentation is a core technical step that supports computer-aided diagnosis of breast diseases and accurate lesion measurement. However, existing segmentation models trained on single-source datasets often suffer from inter-domain distribution shift caused by differences in scanning equipment and scanning protocols across medical institutions, as well as the inherent appearance heterogeneity of lesions. This shift ultimately leads to significant performance degradation. To address this issue, this paper proposes Uncertainty-Calibrated Test-Time Adaptation (UC-TTA), a test-time adaptation framework for cross-dataset breast ultrasound segmentation. The source model is trained on the labeled public BUSI dataset, and only uses the validation set of this dataset to complete calibration. This study conducted verification on three public external datasets: BUS-UCLM, BrEaST-Lesions-USG, and BUS-BRA. The fixed-threshold UC-TTA reduced the expected calibration error from 0.088 to 0.070 on BUS-UCLM, but was accompanied by a small drop in the Dice coefficient, which fell from 0.186 to 0.176. The protected UC-TTA, which integrates threshold calibration on the source domain validation set, achieved a Dice coefficient of 0.389 on this dataset, while reducing the false positive rate for normal images from 0.603 to 0.401. Experiments on ResNet34-U-Net pre-trained on three subsets showed that the source domain calibrated threshold method reached a Dice coefficient of 0.729 on BUS-UCLM, while the protected UC-TTA reached 0.726. The calibrated UC-TTA also increased the Dice coefficient of Attention U-Net from 0.497 to 0.593 on BrEaST-Lesions-USG.