Stromal Tissue Segmentation in Multi-Stained Serial Histopathological Sections of Pancreatic Tumors

In this work we propose and compare different deep learning algorithms for the segmentation of stromal regions in pancreatic histopathological image using three consecutive tissue sections, each uniquely stained with Hematoxylin and Eosin (H&E), Masson’s Trichrome, and Alcian Blue. After a non-rigid registration process, variations in tissue distribution between consecutive slides still persist, which leads to distinct desired segmentations of tissues for each stain, thus underscoring the need for a specific segmentation and co-segmentation approaches to achieve higher accuracy. We compare single stain models, with respect to multi-stain techniques that either consider the multiple stains all at once in training or are based on multi-branch siamese and co-segmentation techniques. We demonstrate superior performance in identifying stromal regions with the multi-stain approaches in comparison to the segmentation techniques applied to individual stains, by effectively utilizing the complementary information each staining technique provides. This advancement is poised to enhance the further evaluation of tumor microenvironment and stromal characteristics in patients with pancreatic cancer.