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

Abstract Accurate segmentation of alopecia areata (AA) lesions is crucial for the diagnosing and staging this non-scarring autoimmune disease, marked by patchy hair loss and psychological impact. However, reliable segmentation remains challenging due to limited annotated datasets, uneven hair distribution, vague lesion boundaries, variability in scalp and hair color, and low quality of photographic images. 1) A novel multi-task network that incorporates a shared segmentation subnetwork and a structure-preserved data augmentation (SDA) subnetwork is proposed. DAGC enhances contextual aggregation, and SDA exploits unlabeled data to improve diversity and robustness. 2) A hierarchical prototype alignment (HPA) module is presented to match graph prototype to the accumulated class-level prototype bank, while simultaneously aligning unlabeled features through contrastive learning. 3) A hierarchical regularization via class-wise voting histogram (HRCVH) strategy is presented by dividing pseudo labels into reliable, fuzzy and unreliable sets with class-wise voting histogram, which is computed by summing the predicted probability scores from the unlabeled image and its perturbed images. Extensive experiments on an in-house AA dataset, the publicly available ISIC dataset and a large-scale CelebAMask-HQ dataset demonstrate that our method significantly outperforms state-of-the-art methods in semi-supervised scenarios. This work introduces a novel framework designed to enhance semantic consistency and facilitate efficient representation learning, particularly in challenging semi-supervised and uncertainty-aware scenarios. In clinical routines, this study is beneficial for quantitative diagnosis, treatment and staging of alopecia areata, which profoundly impacts individual quality of life and mental well-being.