Artificial intelligence improves neoplasia detection in inflammatory bowel disease patients: a pilot study evaluating the added value of a novel dedicated CADe-IBD algorithm for expert and non-expert endoscopists

Aims Dysplasia detection in patients with Inflammatory Bowel Disease (IBD) can be challenging due to the flat and subtle nature of these lesions and background inflammatory / regenerative changes. The miss rate of neoplasia in IBD colon is not truly known. We are entering an artificial intelligence (AI) era and Computer Aided Detection (CADe) may prove its worth in IBD patients. The aim of this study was to evaluate the impact of a novel CADe-IBD algorithm on endoscopists performance

Methods A novel CADe-IBD algorithm was developed using 310,435 frames from IBD colon with sequential training, testing and external validation. High-definition white light (HDWL) videos were prospectively collected during IBD surveillance colonoscopy and edited to short clips with and without lesions. 4 expert endoscopists (those regularly carrying out IBD surveillance including the use of chromoendoscopy) and 4 non-expert endoscopists were asked to review the edited videos for the presence of lesions. Non-experts were then asked to re-watch videos, with the addition of novel CADe-IBD. The videos were randomly re-distributed to reduce recall bias. Ground truth was histology and external review by 3 independent endoscopists.

Results A total of 79 videos (39 with lesions and 40 without lesions) were included. Of the videos with lesions, 46.2% (n=18) were neoplastic and 53.8% (n=21) non-neoplastic. 84.6% (n=33) were non-polypoid lesions (Paris Classification IIa/IIb/IIc) and 15.4% (n=6) were polypoid (Is). 56.4% (n=22) were diminutive (<5mm) in size.

The standalone sensitivity and specificity for lesion detection of CADe-IBD was 84.6% and 82.5% respectively. In comparison, expert sensitivity and specificity was 81.4% and 77.5% and non-experts 76.3% and 63.1% respectively. With the assistance of CADe-IBD, non-expert sensitivity was significantly increased to 92.3% (p<0.001) and specificity to 66.9%. Similar improvements were seen on sub-analysis, including for non-polypoid lesion morphology where the addition of CADe-IBD improved non-expert sensitivity from 77.3% to 93.0%, compared to expert sensitivity of 81.3%.

Conclusions Neoplasia detection in IBD remains challenging regardless of experience. Our data demonstrates that the stand-alone performance of the novel CADe-IBD is as good, if not better, than that of expert endoscopists. However, CADe-IBD could significantly enhance the performance of less experienced endoscopists during colitis surveillance with simple HDWL, without the need for image enhancement technologies or dye spray. This calls for a head-to-head comparison of CADe-IBD with conventional surveillance techniques.

Publication History

Article published online:
15 April 2024