# Pediatric Inflammatory Bowel Disease Tissue Classification From Pathology Slide Images: Detecting Phenotypes Using Computer Vision

## TL;DR

This study explores using computer vision to classify pediatric inflammatory bowel disease tissue from biopsy images, aiming to improve diagnosis and treatment.

## Contribution

The novel contribution is demonstrating that AI can accurately detect abnormal tissue and inflammation in pediatric IBD using weakly supervised learning.

## Key findings

- AI models achieved high accuracy (0.84-0.86) in classifying normal vs abnormal tissue and detecting inflammation and chronic changes.
- The models showed strong performance metrics, including AUC-ROC scores above 0.91 and F1-scores around 0.76-0.79.
- Results suggest that AI can extract interpretable signals from histopathology slides using multiple instance learning.

## Abstract

With the advent of computer vision algorithms, we hypothesize that histopathology images from endoscopic biopsies may be utilized for automated classification of histologic phenotypes, thus guiding Crohn’s disease and ulcerative colitis diagnosis and treatment. The aim of our study is to assess whether artificial intelligence can be used to improve pediatric inflammatory bowel disease outcomes by aiding pathologists with accurate detection of abnormal tissue sections.

Three two-dimensional (2D) convolutional neural networks with multiple instance learning were developed to classify histopathology tissue sections as normal vs abnormal and as containing active inflammation and/or chronic changes/architectural distortion.

The abnormal vs normal classification model achieved an accuracy of 0.84, an area under the receiver operating characteristic curve (AUC-ROC) of 0.91, and an F1-score of 0.79. Precision, sensitivity, and specificity were 0.85, 0.74, and 0.91, respectively. The accuracy for predicting active inflammation was 0.85, AUC-ROC was 0.92, and F1-score was 0.78. The accuracy for predicting chronic changes/architectural distortion was 0.86, with an AUC-ROC of 0.93 and an F1-score of 0.76. All 3 models achieved a Matthews correlation coefficient of 0.67.

The findings resulting from this study are significant primarily because they indicate that there is a strong artificial intelligence–interpretable signal present in endoscopic whole slide imaging, even with the necessary, weakly supervised method of multiple instance learning.