Deep Learning-based Computational Pathology Predicts Origins for Cancers of Unknown Primary
share
Cancers of unknown primary (CUP), represent 1-3 enigmatic disease where the primary site of origin cannot be determined. Pathologists routinely attempt to decipher the site of origin using extensive immunohistochemical staining, however, the rate of identification remains low. Modern cancer therapeutics including immune checkpoint inhibitor therapies are specific to the primary tumor hence broad-spectrum chemotherapy remains the standard-of-care for CUP patients<cit.>. Recent work has focused on using genomics and transcriptomics for identification of tumors of unknown primary. However, genomic testing lacks clinical penetration in low resource settings and may not be available at scale for every patient. Here we show that deep learning-based weakly supervised and multi-task computational pathology on H E slides can be used to determine the metastatic status and site of origin for tumors of unknown primary. We used 17,486 gigapixel whole slide images spread over 18 different organs for training a multi-task deep model to simultaneously identify primary or metastatic status and site of origin. We tested our model on an internal test set of 4,932 cases and achieved a top-1 accuracy of 0.84, a top-3 accuracy of 0.94 and our external test set of 662 cases from 202 different hospitals achieved a top-1 and top-3 accuracy of 0.79 and 0.93 respectively. We further curated a CUP dataset of 717 cases form 202 different medical centers and identified a subset of 290 cases for which a differential diagnosis was assigned. Our model predictions resulted in concordance for 50 chance) and a top-3 agreement of 75 diagnostic mechanism in contrast with or in lieu of immunohistochemical staining and may also reduce the occurrence of CUP cases.
READ FULL TEXT