Abstract
ABSTRACTIn many cancers programmed cell death ligand 1 (PDL1) expression serves as a biomarker for immunotherapies, but its quantification relies on immunohistochemistry (IHC) and few underlying histopathological patterns are established. Digital pathology, combined with deep learning, can augment histopathological assessment and reveal patterns associated with molecular changes, but the presence of heterogeneity within histopathological images, the scale of billions of pixels and the difficulty in acquiring spatially resolved annotations pose challenges for accurate analysis. Here, we present a weakly supervised learning approach using only slide-level supervision for PDL1 expression prediction based on hematoxylin and eosin (H&E) slides. Our methods, MILTS, extends multiple instance learning paradigm (MIL) with the teacher-student framework (TS), which takes the intra-slide heterogeneity into account by dynamically assigning pseudo-labels to different slide regions and retrieves large amounts of unlabeled instances by distillation of the temporal ensemble model. The approach is evaluated on 9,744 tissue slide images across 20 types of solid tumors from TCGA and CPTAC. Among 9 tumors for PDL1 expression serves as an established biomarker, MILTS achieved a weighted average area under curve of 0.83. Predicted patterns within each slide provide insights into the heterogeneity and help identify morphotypes relevant with PDL1 expression, which include mixed inflammatory stroma with relatively high abundance of eosinophils and a cribriform growth pattern of tumor cells. This study provides a new algorithm for predicting molecular changes from H&E images and provides histological links of PDL1 expression, and thus demonstrates the potential of deep learning in discovering diverse histological patterns.
Publisher
Cold Spring Harbor Laboratory