The consistency regularization method is a widely used semi-supervised method that uses regularization terms constructed from unlabeled data to improve model performance. Poor-quality target predictions in regularization terms produce noisy gradient flows during training, resulting in a degradation in model performance. Recent semi-supervised methods usually filter out low-confidence target predictions to alleviate this problem, but also prevent the model from learning features from unlabeled data in low-confidence regions. Specifically, in medical imaging and other cross-domain scenarios, models are prone to producing large numbers of low-confidence predictions. To improve the quality of target predictions while utilizing unlabeled data more efficiently, we propose an uncertainty-aware semi-supervised method that incorporates the breast anatomical prior, for pectoral muscle segmentation. Our method has a typical teacher-student dual model structure, where uncertainty is used to distinguish between high- and low-confidence predictions in the teacher model output. A low-confidence prediction refinement module was designed to refine the low-confidence predictions by incorporating high-confidence predictions and a learned anatomical prior. The anatomical prior, as regularization of the target predictions, was learned from annotations and an auxiliary task. The final target predictions are a combination of high-confidence teacher predictions and refined low-confidence predictions. The proposed method was evaluated on a dataset containing 635 data points from three data centers. Compared with the baseline method, the proposed method showed an average improvement in DICE index of 1.76, an average reduction in IoU index of 3.21, and an average reduction in HD index of 5.48. The experimental results show that our method generalizes well to the test set and outperforms other methods in all evaluation metrics.
Keywords: deep learning; segmentation; uncertainty.