Artificial Intelligence in Colon Neoplasm Diagnosis: Development, Implementation of Technology, and Initial Results

Abstract

Introduction. Colorectal cancer is a critical issue that requires prompt diagnosis and treatment. According to the WHO and IARC for 2020, about 1.93 million cases of CRC are registered annually in the world. Despite the advancement of endoscopic equipment, there is still a significant number of missed cases of colon cancer after diagnostic colonoscopy, with a range of 2.1 % to 5.9 %. The proportion of such missed cases for precancerous pathology has reached 32.8 %. Multiple factors contribute to missed pathology, one of which is the "human factor". The frequency of detected pathology highly depends on the qualifications and experience of the endoscopist.

Aim. To evaluate the effectiveness of diagnostic colonoscopy using an artificial intelligence system in detecting colon neoplasms.

Materials and methods. From 2021 to 2022, the Chelyabinsk Regional Clinical Center of Oncology and Nuclear Medicine together with the Russian company EVA Lab (EVA Lab LLC), developed, tested, and implemented a medical decision support system (DMSS) based on artificial intelligence algorithms. The study comprised an analysis of 944 patients with various pathologies of the large intestine, including 338 men (41.1 %) and 556 women (58.9 %). The mean age of men and women was 64 ± 12.9 years and 63 ± 10.2 years, respectively. All patients were categorized into two groups: a retrospective control group comprising 634 patients formed before the implementation of the AI-based system in 2020, and a prospective cohort comprising 310 patients formed from 2020 (the time of implementation of the AI-based system). In both groups, diagnostic colonoscopies were performed by the same endoscopists with a minimum of 10 years of experience.

Results. Among the control group, 358 (56.4 %) cases of colon neoplasms were detected, while in the study group, 204 (65.8 %) cases were detected. Notably, the highest efficiency of the AI system was observed in detecting neoplasms up to 1.0 cm in diameter. When comparing the frequency of neoplasm detection in patient categories with sizes up to 0.5 cm and 0.5 to 1.0 cm, a statistically significant difference of 15.7 % was observed, with neoplasms being detected more frequently in the study group than in the control group (p<0.001). However, with neoplasm sizes over 1.0 cm in diameter, no significant differences were found between the two groups.  Biopsy was performed 13 % more frequently in the study group.

Conclusion. The AI system effectively detected neoplasms of any size with 80.7% sensitivity.  The system showed a 13.7 % higher probability of detecting neoplasms less than 1.0 cm in diameter and a 9.7 % higher probability of detecting tubular adenomas of all sizes.