Use of artificial intelligence to measure colorectal polyp size without a reference object

Abstract Polyp size is crucial for determining colonoscopy surveillance intervals. We present an artificial intelligence (AI) model for colorectal polyp size measurement without a reference object. The regression model for polyp size estimation was developed using outputs from two SegFormer models, segmentation and depth estimation. Initially built on colonoscopic images of polyp phantoms, the model underwent transfer learning with 1,304 real-world images. Testing was conducted on 178 images from 52 polyps, independent of the training set, using a snare as the ground truth for size comparison with the AI-based model. Polyps were classified into three size groups: ≤ 5 mm, 5–10 mm, and ≥ 10 mm. Error rates were calculated to evaluate discrepancies in actual size values between the AI model and the snare method. Precision indicated the positive predictive value per size group and recall and Bland-Altman were also conducted. The Bland-Altman analysis showed a mean bias of –0.03 mm between methods, with limits of agreement from –1.654 mm to 1.596 mm. AI model error rates for actual size discrepancies were 10.74%, 12.36%, and 9.89% for the ≤ 5 mm, 5–10 mm, and ≥ 10 mm groups, respectively, averaging 11.47%. Precision values were 0.870, 0.911, and 0.857, with overall recall of 0.846. Our study shows that colorectal polyp size measurement by AI model is practical and clinically useful, exhibiting low error rates and high precision. AI shows promise as an accurate tool for measurement without the need for a reference object during screening colonoscopy.