Deployment of an Artificial Intelligence Histology Tool to Aid Qualitative Assessment of Histopathology Using the Nancy Histopathology Index in Ulcerative Colitis
David T. Rubin, Olga Kubassova, Christopher R. Weber, Shashi Adsul, Marcelo Freire, Luc Biedermann, Viktor H. Koelzer, Brian Bressler, Wei Xiong, Jan Hendrik Niess, Matthias S. Matter, Uri Kopylov, Iris Barshack, Chen Mayer, Fernando Magro, Fátima Carneiro, Nitsan Maharshak, Ariel Greenberg, Simon P. Hart, Jamshid Dehmeshki, Laurent Peyrin‐Biroulet
Abstract
BACKGROUND: Ulcerative colitis (UC) is a chronic inflammatory bowel disease characterized by increased stool frequency, rectal bleeding, and urgency. To streamline the quantitative assessment of histopathology using the Nancy Index in UC patients, we developed a novel artificial intelligence (AI) tool based on deep learning and tested it in a proof-of-concept trial. In this study, we report the performance of a modified version of the AI tool. METHODS: Nine sites from 6 countries were included. Patients were aged ≥18 years and had UC. Slides were prepared with hematoxylin and eosin staining. A total of 791 images were divided into 2 groups: 630 for training the tool and 161 for testing vs expert histopathologist assessment. The refined AI histology tool utilized a 4-neural network structure to characterize images into a series of cell and tissue type combinations and locations, and then 1 classifier module assigned a Nancy Index score. RESULTS: In comparison with the proof-of-concept tool, each feature demonstrated an improvement in accuracy. Confusion matrix analysis demonstrated an 80% correlation between predicted and true labels for Nancy scores of 0 or 4; a 96% correlation for a true score of 0 being predicted as 0 or 1; and a 100% correlation for a true score of 2 being predicted as 2 or 3. The Nancy metric (which evaluated Nancy Index prediction) was 74.9% compared with 72.3% for the proof-of-concept model. CONCLUSIONS: We have developed a modified AI histology tool in UC that correlates highly with histopathologists' assessments and suggests promising potential for its clinical application.