Fusing Domains for Tumor Detection in Frozen Section Histopathology Using Joint Feature Alignment and Self-Knowledge Distillation

Robustness in neural networks is of paramount importance in the field of image classification. However, challenges such as misclassifications arising from variations in image characteristics persist. Consequently, researchers are actively seeking solutions to address domain shift challenges and enhance classification performance. Domain shift, which is particularly prevalent in whole-slide image (WSI) data within the context of histopathology, is exacerbated by differences in WSI scanners and the status of WSIs. This study focuses on the domain shift encountered in tumor detection in breast lymph nodes by leveraging the publicly available Camelyon16 dataset and an in-house dataset. To address this issue, we propose a simple yet effective self-knowledge distillation approach with contrastive embedding alignment across domains. By transferring knowledge from the source data to the target data, the experimental results demonstrated remarkable adjustments in the domain shift, providing promising insights for bolstering the robustness of WSI-based image classification systems. These findings have significant implications for the advancement of tumor detection accuracy and reliability of histopathological WSI analysis.