Category-specific unlabeled data risk minimization for ultrasound semi-supervised segmentation

Achieving accurate computer-aided analysis of ultrasound images is challenging, since not only its image artifacts but also the difficulties in collecting large-scale pixel-wise annotations from experts for training. Semi-supervised segmentation is a solution for learning from labeled and unlabeled data, which mainly focuses on generating pseudo annotations for unlabeled data or learning consistent features in enhanced views of images to enhance model generalization. However, anatomically, diverse learning difficulties across tissues are overlooked, and, technically, the estimation and minimization of empirical risk for unlabeled training data are largely ignored. Motivated by them, this work proposes a semi-supervised segmentation model, named CSUDRM, with two modules. The former is called category-specific distribution alignment (CSDA), which learns more consistent feature representations of the same class across labeled and unlabeled data. Moreover, it enhances feature space intra-class compactness and inter-class discrepancy and provides category-specific penalties for more robust learning. The latter one is Unlabeled Data Risk Minimization (UDRM). It minimizes the risk on the entire training data, which distinguishes it from most existing works that merely optimize labels. The risk of unlabeled data is estimated by a novel learnable class prior estimator, with the help of distributional hints from CSDA. This design could reinforce the robustness of the model and achieve stable segmentation. CSUDRM achieves state-of-the-art performances on four ultrasound datasets. Extensive ablation studies, including quantitative comparisons, feature space visualization, and robustness analysis, demonstrate the superiority of our designs.