Unmasking the tissue microecology of ductal carcinoma in situ with deep learning

Abstract

Despite increasing evidence supporting the clinical relevance of tumour infiltrating lymphocytes (TILs) in invasive breast cancer, TIL spatial distribution pattern surrounding ductal carcinoma in situ (DCIS) and its association with progression is not well understood. To characterize the tissue microecology of DCIS, we designed and tested a new deep learning pipeline, UNMaSk (UNet-IM-Net-SCCNN), for the automated detection and simultaneous segmentation of DCIS ducts. This new method achieved the highest sensitivity and recall over cutting-edge deep learning networks in three patient cohorts, as well as the highest concordance with DCIS identification based on CK5 staining. Following automated DCIS detection, spatial tessellation centred at each DCIS duct created the boundary in which local ecology can be studied. Single cell identification and classification was performed with an existing deep learning method to map the distribution of TILs. In a dataset comprising grade 2-3 pure DCIS and DCIS adjacent to invasive cancer (adjacent DCIS), we found that pure DCIS cases had more TILs compared to adjacent DCIS. However, TILs co-localise significantly less with DCIS ducts in pure DCIS compared with adjacent DCIS, suggesting a more inflamed tissue ecology local to adjacent DCIS cases. Our experiments demonstrate that technological developments in deep convolutional neural networks and digital pathology can enable us to automate the identification of DCIS as well as to quantify the spatial relationship with TILs, providing a new way to study immune response and identify new markers of progression, thereby improving clinical management.