AI in Predicting Polyp Pathology and Endoscopic Classification

Background: Colonoscopy with optical diagnosis based on the appearance of polyps can guide the selection of endoscopic treatment methods, reduce unnecessary polypectomy procedures and the need for tissue pathological diagnosis, and formulate follow-up strategies in a timely manner \[1\]. This approach significantly alleviates the economic burden on patients and the healthcare system and can effectively ease the tension on clinical resources \[2\]. Various endoscopic polyp classification methods, including Pit Pattern \[3\], NICE \[4\], WASP \[5\], and MS \[6\], are used to determine pathological types. However, mastering these classification methods requires endoscopists to undergo extensive training, and due to the inherent flaws in each method, no single endoscopic classification method can accurately diagnose all types of polyps to meet the requirements of optical diagnosis. This limitation has hindered the widespread application of optical diagnosis in clinical practice \[7\]. The application of artificial intelligence technology in this field, known as computer-aided diagnosis (CADx), has seen rapid development in recent years. Numerous large-scale, prospective studies have demonstrated that the accuracy of CADx technology for optical diagnosis of minute lesions (\<5mm) has essentially met the threshold set by European and American endoscopy societies for optical diagnosis \[8,9\]. However, the diagnostic efficacy of CADx for polyps ≥5mm remains unclear. Moreover, current research is mostly limited to distinguishing between common adenomas and hyperplastic polyps, with little attention given to serrated lesions, which are also precancerous lesions and progress even more rapidly, and are more challenging for endoscopists to assess. These reasons prevent CADx from being widely applied in clinical practice for real-time accurate judgment of polyp pathological types.