Hierarchical prototype alignment and regularization for semi-supervised alopecia areata segmentation

The study introduces a novel framework for improving the segmentation of alopecia areata (AA) lesions, which is crucial for diagnosing and staging this autoimmune disease characterized by patchy hair loss. The proposed method includes a multi-task network with a shared segmentation subnetwork and a structure-preserved data augmentation (SDA) subnetwork, enhancing contextual aggregation and exploiting unlabeled data for improved robustness. A hierarchical prototype alignment (HPA) module and a hierarchical regularization strategy via class-wise voting histogram (HRCVH) are also presented. Extensive experiments on various datasets demonstrate that this method significantly outperforms existing state-of-the-art methods in semi-supervised scenarios, aiding in the quantitative diagnosis and treatment of AA, which impacts quality of life and mental well-being.