Domain-Centroid-Guided Progressive Teacher-Based Knowledge Distillation for Source-Free Domain Adaptation of Histopathological Images

Deep neural networks are commonly used for histopathology image analysis. However, such data-driven models are sensitive to style variances across scanners and suffer a significant performance degradation as a result. Although the network performance can be improved by using domain adaptation methods, the source dataset required to perform the adaptation process is generally unavailable. This study shows that the performance degradation of deep neural networks when applied to histopathology images is the result partly of the wide distribution of the features generated when inferring the features of the target model using the feature centers of the source model. To address this problem, a teacher-student framework, designated as domain-centroid-guided progressive teacher-based knowledge distillation (DCGP-KD), is proposed which aims to learn compact target features in order to provide more accurate pseudo labels for the target model without the need for the original source dataset. In the proposed framework, the class-wise feature centers of the source data are progressively adapted to the distribution of the target data, and compact target features are then generated by gathering the features based on their class-wise centers. A strategy is additionally proposed to prevent catastrophic forgetting during the progressive adaption process. Finally, a prediction consistency loss function is introduced to improve the robustness of the target dataset. The feasibility of the proposed framework is demonstrated experimentally for the illustrative case of the tumor classification of histopathological images with staining variations. The results show that DCGP-KD provides a promising assistive tool for pathologists in various histopathological analysis tasks.