Artificial intelligence-assisted diagnosis of rectal neuroendocrine tumors during white-light endoscopy.

BACKGROUND

Due to their significantly lower incidence than colorectal polyps and macroscopic features resembling those of hyperplastic polyps, rectal neuroendocrine tumors (rNETs) are frequently misdiagnosed and resected as polyps. To date, no reports have been written on the application of artificial intelligence for assisting in the white-light endoscopy of rNETs.

AIM: To establish a neuroendocrine tumor lesion detection algorithm based on the YOLOv7 model and evaluate the performance of the algorithm in identifying neuroendocrine tumors.

METHODS

In total, 137748 white-light endoscopic images were collected in this study, including 2232 images of rNET, 4429 images of submucosal lesions other than rNET, 42563 images of polyps, and 88593 images of normal mucosa. All the images were randomly divided into a training set, a validation set, and a test set. To evaluate the ability of the algorithm to diagnose rNETs, we selected 1578 images to form the test set. The performance of the algorithm was compared with that of endoscopists at different levels.

RESULTS

The accuracy of the algorithm in identifying rNET from all the images was 97.8%, the sensitivity was 72.6%, the specificity was 99.7%, the positive predictive value was 93.9%, and the negative predictive value was 98.1%.

CONCLUSION

Our model, which was based on YOLOv7, could effectively detect rNET lesions, which was better than that of most endoscopists.