Category-Level Regularized Unlabeled-to-Labeled Learning for Semi-supervised Prostate Segmentation with Multi-site Unlabeled Data

Segmenting prostate from MRI is crucial for diagnosis and treatment planning of prostate cancer. Given the scarcity of labeled data in medical imaging, semi-supervised learning (SSL) presents an attractive option as it can utilize both limited labeled data and abundant unlabeled data. However, if the local center has limited image collection capability, there may also not be enough unlabeled data for semi-supervised learning to be effective. To overcome this issue, other partner centers can be consulted to help enrich the pool of unlabeled images, but this can result in data heterogeneity, which could hinder SSL that functions under the assumption of consistent data distribution. Tailoring for this important yet under-explored scenario, this work presents a novel Category-level regularized Unlabeled-to-Labeled (CU2L) learning framework for semi-supervised prostate segmentation with multi-site unlabeled MRI data. Specifically, CU2L is built upon the teacher-student architecture with the following tailored learning processes: (i) local pseudo-label learning for reinforcing confirmation of the data distribution of the local center; (ii) category-level regularized non-parametric unlabeled-to-labeled learning for robustly mining shared information by using the limited expert labels to regularize the intra-class features across centers to be discriminative and generalized; (iii) stability learning under perturbations to further enhance robustness to heterogeneity. Our method is evaluated on prostate MRI data from six different clinical centers and shows superior performance compared to other semi-supervised methods.