Automated Lewis Score Estimation for Small Bowel Crohn Disease Using Capsule Endoscopy: Retrospective Cohort Study

Small bowel capsule endoscopy (SBCE) is widely used to evaluate small bowel involvement in Crohn disease (CD); however, manual Lewis score calculation is time-consuming and susceptible to interobserver variability.

This study aimed to develop and preliminarily validate a multitask deep learning system for automated Lewis score estimation in patients with small bowel CD undergoing SBCE.

We developed Eval-Model, a multitask framework integrating You Only Look Once (YOLO)-style lesion detection and localization for ulcers and edema, a stenosis classification branch, and a long short-term memory-based temporal regression module. The primary dataset comprised SBCE examinations from 814 patients at the Children's Hospital of Fudan University. In this dataset, 725 (89.1%) cases were included for lesion-level annotation, and 89 (10.9%) cases with expert-consensus Lewis scores were used for scoring model development and internal testing. These 89 cases comprised 49 (55.1%) training cases and 40 (44.9%) internal test cases. All splits were performed at the patient level. A retrospective same-center temporal validation cohort of 20 consecutive cases from 2026 and 5 control cases were used for preliminary assessments of temporal robustness and specificity.

In the 40-case internal Lewis score test set, Eval-Model achieved a mean squared error of 14.693 (95% CI 8.72-22.15), root mean squared error of 3.833 (95% CI 2.95-4.71), mean absolute error of 2.9 (95% CI 2.1-3.8), R² of 0.82 (95% CI 0.70-0.90), and explained variance of 0.83 (95% CI 0.72-0.91). For lesion recognition, the model achieved an overall mean average precision of 0.94 with an inference time of 15 ms per frame. In the 20-case same-center temporal validation cohort, the mean squared error was 26.848 (95% CI 12.51-45.33), the root mean squared error was 5.182 (95% CI 3.54-6.73), the mean absolute error was 3.3 (95% CI 2.2-4.6), R² rounded to 1.00 (95% CI 0.98-1.00), and the explained variance rounded to 1.00 (95% CI 0.98-1.00). These temporal estimates require cautious interpretation because of the small, same-center cohort.

Eval-Model demonstrated preliminary feasibility for automated Lewis score estimation from SBCE in small bowel CD. The system may serve as a standardized adjunct to expert review, but larger multicenter prospective validation with independent external annotation is required before clinical deployment.