Deep learning-based quantification of collagen and associated features from H&E-stained whole slide pathology images across cancer types

Collagen, a major component of the extracellular matrix, undergoes reorganization during tumorigenesis. However, visualizing collagen in tissue currently requires special stains (complicated by variability) or imaging (hindered by reliability, cost, and throughput). Here, we describe an approach to exhaustively detect and quantify collagen in hematoxylin and eosin (H&E)-stained images. Polarization images of picrosirius red-stained slides served as ground-truth training images for development of inferred quantitative multimodal anisotropy imaging (iQMAI), a deep learning model that infers collagen directly from H&E-stained whole-slide images. Individual collagen fibers were extracted after iQMAI inference, and features describing collagen intensity and fiber morphology were computed. iQMAI outputs were measured against polarization imaging and were compared to tissue composition, gene expression, and survival in lung adenocarcinoma, lung squamous cell carcinoma, hepatocellular carcinoma, and pancreatic adenocarcinoma datasets from the cancer genome atlas. iQMAI-derived collagen predictions and fiber features (fiber tortuosity, length, width, and relative angle) are correlated with polarization-based measurements. In pancreatic adenocarcinoma, fiber density and width are negatively associated with the LRRC-15 gene expression signature, and increased fiber width is associated with longer overall survival. iQMAI, a deep learning model identifying collagen in H&E-stained images, allows exhaustive, spatially-resolved quantification of collagen morphology, enabling investigation of the interplay between collagen and the tumor microenvironment. The association of collagen features with immunosuppressive fibroblasts and outcome in pancreatic adenocarcinoma demonstrates the potential of iQMAI. Understanding the relationship between collagen, tumor composition, and disease progression may aid the development of effective oncology therapies. Collagen is a structural protein found throughout the body. In diseases, including cancer, the amount and organization of collagen is altered. However, evaluating collagen in tumor pathology samples has required the use of special stains or imaging approaches, which can be variable and expensive, limiting widespread adoption. We developed a computational model (termed iQMAI) to predict and quantify collagen in images from routinely-stained tissue samples. This tool yields information about the amount of collagen present and the organization of collagen fibers. The collagen distribution measured by iQMAI in pancreatic cancer samples was found to be associated with poor outcomes. The use of this model has the potential to identify collagen patterns associated with outcomes for patients with cancer. Nguyen, Zhang et al. detail a deep learning approach to infer collagen from digitized H&E-stained whole slide images. In pancreatic adenocarcinoma, outputs of this model are associated with an immunosuppressive gene expression signature and overall survival.