DermNet: An Incremental Knowledge Distillation Framework for Accurate and Robust Skin Layer Segmentation

Background Accurate segmentation of skin layers in histopathological whole-slide images (WSIs) is important for identifying disease-associated structural changes and supporting computer-aided dermatopathological analysis. However, existing approaches are often restricted to single magnification levels and may inadequately capture the heterogeneous morphological characteristics of individual skin layers. We propose DermNet, a multi-resolution class-incremental learning framework designed to achieve accurate skin layer segmentation while mitigating catastrophic forgetting. Methods DermNet employs a SegFormer backbone within a teacher-student architecture and progressively learns class-specific features from WSIs acquired at 10x, 5x, and 2x magnifications. The framework incorporates exemplar replay, mutual knowledge, distillation, classification, Kullback-Leibler divergence losses, boundary loss, and Elastic Weight Consolidation to preserve previously learned representations while incorporating new tissue classes. The model was trained and evaluated using a publicly available dataset of 290 WSIs, with segmentation focused on the epidermis, papillary dermis, reticular dermis, hypodermis, keratin, and background. Results DermNet consistently outperformed U-Net and alternative SegFormer training strategies across all three magnification levels. At 10x magnification, DermNet achieved Dice scores of 0.95 for epidermis, 0.92 for papillary dermis, 0.94 for reticular dermis, 0.97 for keratin, and 0.98 for hypodermis, outperforming UNet++ and DeepLabv3+. TransUNet, SwinUNet, and nnU-Net under the same experimental conditions. Evaluation on an independent publicly available dataset further demonstrated cross-dataset generalizability, with Dice scores ranging from 0.89 to 0.97 across classes. Conclusion DermNet enables accurate and robust multi-resolution skin layer segmentation by combining class-incremental learning with knowledge distillation and mechanisms for preserving previously learned features. Its performance across magnification levels and an independent dataset supports its potential for scalable and generalizable automated dermatopathological image analysis.