Exploring Domain Generalization in Semantic Segmentation for Digital Histopathology: A Comparative Evaluation of Deep Learning Models

With the rise in clinicopathologic prognosis and the introduction of whole slide imaging scanners, artificial intelligence systems are being suggested to assist pathologists in their decision-making processes. Despite advancements, the variability in digital pathology images, due to different organs, tissue preparation methods, and image acquisition processes, poses significant challenges. This variability, known as domain shift, affects the performance of machine learning models trained on specific datasets. Addressing these challenges, this paper explores domain generalization (DG) in semantic segmentation for digital histopathology images. We systematically evaluate the DG capabilities of six prominent deep learning models across two novel adenocarcinoma segmentation datasets. Our comparative analysis provides insights into the models' effectiveness in mitigating domain shift, contributing to the advancement of DG in computational pathology.