AI in IBD - Assessing Histological Disease Activity
Aim
Ulcerative colitis (UC) is an idiopathic inflammatory bowel disease (IBD) that may dramatically impact patient quality of life. Introduction in the last decade of a range of biological drugs has reduced the disease burden and may allow us to define histologic remission as an ambitious new treatment target. This has stimulated interest in the development of histological scoring systems to assess disease chronicity and activity. The partially validated DCA score was specifically designed for use in daily clinical practice. This scoring system includes an assessment of the severity of disease chronicity (C), activity (A), and extent of the histologic changes (D). One hurdle for widespread implementation into clinical practice is the reported inter-observer variability.
To tackle the inter-observer variability, we aim to develop a computer algorithm to automatically perform the score on scans of whole slide images. The hematoxylin-eosin (H&E) stained slides from a retrospective cohort of adult patients with UC who received an endoscopy with biopsies at the Inselspital Bern during disease follow-up between 2011 and 2021 will be scanned. Two expert gastrointestinal pathologists blinded to the clinical data will independently perform the DCA-score. Afterward, discrepancies will be discussed in a consensus meeting to determine the consensus DCA score.
The core computer algorithm involves using AlexNet. The pre-trained model vectors will be open-source data trained on the “Image Net” classification dataset. Moreover, some functions from the open-source library “OpenCV” will be incorporated. The scanned slides will be randomly split into train and test sets with an 80:20 ratio. The only available labels for the algorithm will be the consensus DCA score. To enrich our dataset, small segments of the scans will be taken at various magnifications and various image processing transformations shall be applied (including but not limited to Gaussian Blur, Rotation, and Flipping) to make the model more robust to differences in scanning devices and histological slide preparation. While the model ranks individual small fragments of the image, the overall slide score shall be determined by an aggregate vote based on all the fragment rankings. Afterward, scans from the first available follow-up endoscopy will constitute a validation set, which will serve as a proxy of how the algorithm performs in a real-world setting.
The Model’s Gradcam Heatmap shows localization of a Neutrophil, indicating Active Inflammation.
Methods
Histological disease activity in ulcerative colitis is captured by the DCA score, which rates three features — Distribution, Chronicity and Activity — each on a three-point scale, with distribution reflecting how much of the biopsy is affected by inflammation. Because these assessments vary considerably between observers, we trained a deep learning model to compute the DCA score automatically. Biopsies from 117 adult ulcerative colitis patients, taken during routine colonoscopy, were used to build the model — 537 H&E-stained whole slide images for training and 299 for testing — with an expert gastrointestinal pathologist's scores serving as the reference standard. The model was trained in a weakly supervised, multiple-instance learning setup, using only slide-level labels without any region-of-interest annotations, and combined features learned at 2.5×, 5× and 10× magnification through a random forest classifier to produce the final score.
Results
The model reproduced the pathologist's scoring well for the coarser distinctions and, tellingly, appeared to "read" the tissue much as a pathologist does — judging architectural distortion at low magnification and then cellular detail, such as neutrophils, at higher magnification (confirmed by GradCam heatmaps). Combining features across magnifications outperformed any single level. Performance on the test set:
| Component | AUC | Normal vs inflamed | Grade 1 vs 2 | Full 3-level |
|---|---|---|---|---|
| Distribution (D) | 0.972 | 90.9% | 81.1% | 85.8% |
| Chronicity (C) | 0.963 | 91.4% | 70.1% | 80.9% |
| Activity (A) | 0.961 | 90.9% | 53.0% | 78.6% |
Across the full 9-class DCA score, overall accuracy was 64.8% (AUC 0.972). The model reliably separated normal from inflamed tissue (~91%), but grading activity finely (grade 1 vs 2) was harder — the multiple-instance approach tended to miss crypt abscesses, the feature that distinguishes those grades.
In short, a weakly supervised model can reproduce expert DCA scoring closely enough to capture the presence and extent of inflammation, pointing towards more consistent, less observer-dependent disease-activity assessment in ulcerative colitis — with finer activity grading the clear target for further work.
Members
Kartik Kohli
Aart Mookhoek
Amjad Khan
Publications
Kohli, K., Khan, A., Müller, F., Cathomas, G., Lugli, A., Zlobec, I. & Mookhoek, A. Automated assessment of histological disease activity in ulcerative colitis using a deep learning algorithm. Journal of Crohn's and Colitis 18(Supplement_1), i753–i754 (2024), abstract P357. https://doi.org/10.1093/ecco-jcc/jjad212.0487