Whole-Slide Image Inference Using Deep Learning

All the slides were scanned at a magnification of 40× using a Leica Aperio AT2 scanner (Leica Biosystems). The resulting whole-slide images were fed to a ResNet-18 neural network,²⁷ which was developed and validated to classify colorectal polyps into 4 classes (ie, tubular adenoma, tubulovillous or villous adenoma, sessile serrated polyp, and hyperplastic polyp) with an independent set of 508 slides from DHMC and was previously validated with 238 external slides from 24 different institutions.^24,28 The model used a sliding-window approach in which predictions were made on patches of 224 × 224 pixels. These predictions were then used to calculate the percentage of patches, a proxy for the percentage of area, attributed to each class in the whole-slide image. The percentage of patches for each class was then used in a decision tree to determine the overall class of the whole-slide image.²⁴ For our digital system, we extracted the percentage of patches attributed to each class, the coordinates for the regions of interest highlighted by the classifier for each class, and the whole-slide image prediction.