Application of artificial intelligence in assessing the size of colon tumors during endoscopic examination

Authors:

Abstract
Introduction. The risk of malignancy of precancerous colonic lesions directly depends on the size of the neoplasms detected during colonoscopy. According to various authors, neoplasms up to 2.0 cm in diameter have a 12.5% risk of harboring malignant cells. If the neoplasm is larger than 2.0 cm, the risk of malignancy increases significantly and, according to various authors, ranges from 31 to 57%. The endoscopist’s subjective assessment of neoplasm size often does not correspond to the actual size, leading to inappropriate treatment decisions. Aim. To develop and evaluate the effectiveness of a hardware-based subsystem for measuring the size of colonic neoplasms (AI-assisted neoplasm size estimation, AIANSE) during colonoscopy with AI-assisted colonoscopy detection (AIAC). Materials and methods. From 2022 to 2024, the Chelyabinsk Regional Clinical Center of Oncology and Nuclear Medicine, in collaboration with the Russian company EVA Lab, developed and implemented a clinical decision support system based on artificial intelligence algorithms. The study included 1,763 patients, from whom approximately 27,000 frames were labeled. Results. Computer processing of endoscopic images demonstrated its effectiveness in both tumor detection and tumor size estimation. This is confirmed by comparing colonoscopy image analysis data with the actual dimensions of the neoplasms determined after surgical treatment. The system possesses self-learning capabilities, as evidenced by the increase in specificity (from 63% to 70%), sensitivity (from 64% to 70%), and accuracy (from 66% to 71%) between 2022 and 2024. Conclusions. AIANSE, a computer-based tumor analysis and size measurement system, was developed for the first time. This system demonstrated its effectiveness in tumor size estimation, with an accuracy of 71%. This development demonstrated its greatest effectiveness in measuring tumor sizes of 5 mm in diameter, with an accuracy of 76%.
Keywords

This publication has 25 references indexed in Scilit: