Deep Learning for Predicting Pediatric Crohn's Disease Using Histopathological Imaging

Abstract

The current gold standard for Crohn's disease diagnosis involves the examination of biopsied tissue by a trained physician. However, endoscopic images and histological features are only evident when the appropriate biopsy site is chosen and the image is of high quality [1]. Thus, to prevent delayed diagnoses or reclassifications over time, additional tools to reinforce pathologists' diagnoses are necessary. Recent studies have showcased successful applications of deep learning for developing whole-slide classifications of digital histology images. In this study, we developed a patch-level image classification model for prediction of Crohn's disease using a convolutional neural network. This study obtained data from two different hospitals: INOVA and Cincinnati Children's Hospital Medical Center (CCHMC). When trained and validated on the same data set, our INOVA and CCHMC models achieved validation accuracies of 84.6 % and 93.9 %, respectively. However, the models performed poorly when trained on data from one site and tested on data from the other site. To investigate this issue, we built an additional patch-level model that was able to predict hospital source of the biopsy with 99 % accuracy. These results suggest the presence of site-specific artifacts which are detectable by machine learning models. We reduced the effects of these artifacts using color-normalization, image cropping, and other transformations, lowering site-predictive accuracy to 74%. Therefore, we suggest further works investigate reasons for inter-site biopsy differences such that site-generalizable, histopathological deep learning models can be developed.