Artificial Intelligence System Detects Polypoid Lesions in Inflammatory Bowel Disease

The “You Only Look Once” computer-aided detection system is capable of detecting polypoid lesions in IBD.

Researchers have developed the first automated system capable of detecting polypoid lesions in inflammatory bowel disease (IBD), according to study results presented at the Advances in Inflammatory Bowel Diseases (AIBD) 2022 conference, held from December 5 to 7, 2022, in Orlando, Florida.

Much effort has been spent on developing artificial intelligence (AI) models that are capable of detecting colorectal polyps in real-time. To date, there has been no such model developed lesions in the setting of IBD.

The You Only Look Once (YOLO) computer-aided detection system (CADe) for detecting polypoid lesions in IBD was developed using a training and retraining protocol. The model was first pretrained with 8000 endoscopic images and trained with 2268 annotated white light colon polyp images from patients without IBD. The model was tested using 2016 unlabeled images of IBD polypoid lesions, and images were categorized into dysplastic, nondysplastic, pseudopolyp, serrated changes (SECs), and sessile serrate adenoma categories. After categorization, 8 clinicians hand-labeled images. Next, 80% of the total IBD polypoid lesion dataset was used to retrain the model, and the remaining 20% was used in iterative testing.

The development of this tool is the first step to the creation of several technologies that are inclusive to patients with IBD…

The original YOLO algorithm that was trained using only non-IBD images was best at correctly identifying dysplastic polyps and identified pseudopolyps and SECs poorest. The poor performance in identifying pseudopolyps and SECs was likely due to the fact that the original training did not include any IBD images.

The overall sensitivity of the original YOLO model was 0.50, positive predictive value (PPV) was 0.97, false positive rate (FPR) was 1.7%, and accuracy was 0.64.

After retraining, (YOLO version 4), the sensitivity to detect all polyp types improved with an overall sensitivity of 0.95, PPV of 0.95, FPR of 5%, and accuracy of 0.95.

A total of 9 lesions were missed by YOLO version 4. Three of these lesions had a Mayo Score of 0, 2 had a score of 1, 2 a score of 2, and 2 a score of 3. These trends indicated that the performance of the YOLO model was likely not affected by background inflammation.

In this study, the first automated algorithm for detecting polypoid lesions, serrated lesions, and pseudopolyps in IBD was developed. The performance of the algorithm did not appear to be affected by background inflammation.

“The development of this tool is the first step to the creation of several technologies that are inclusive to patients with IBD, increasing polyp detection in this patient cohort and aiding the endoscopist in automated detection of polypoid dysplasia,” the study authors noted.

References:

Guerrero Vinsard D, Fetzer J, Agrawal U, et al. Development of an artificial intelligence tool for detection of polypoid lesions in inflammatory bowel disease (IBD-CADe). Abstract presented at: AIBD 2022; December 5-7, 2022; Orlando, FL. Abstract 26.