Attention-Based Whole-Slide Image Compression Achieves Pathologist-Level Prescreening of Multiorgan Routine Histopathology Biopsies.

Abstract

Screening programs for the early detection of cancers, such as colorectal and cervical cancers, have led to an increased demand for histopathological analysis of biopsies. Advanced image analysis with deep learning has shown the potential to automate cancer detection in digital pathology whole-slide images. In particular, weakly supervised learning can achieve whole-slide image classification without the need for tedious, manual annotations, using only slide-level labels. Here, we used data from n = 12,580 whole-slide images from n = 9141 tissue blocks to train and validate a weakly supervised deep learning approach based on Neural Image Compression with Attention (NIC-A) using labels extracted from pathology reports. We also introduced slide packing, a method that merges tissue from multiple slides of the same tissue block into a single "packed" image linked to block-level labels. NIC-A classifies colon and cervical tissue slides into cancer, high-grade dysplasia, low-grade dysplasia, and normal tissue and detects celiac disease in duodenal biopsies. We validated NIC-A for colon and cervix against panels of 4 and 3 pathologists, respectively, on cohorts from 2 European centers. We show that the proposed approach reaches pathologist-level performance in detecting and classifying abnormalities, suggesting its potential to assist pathologists in prescreening workflows by reducing workload in routine digital pathology diagnostics.