Development and validation of a real-time computer-aided measuring system for colorectal polyp size (with video)

Polyp size measurement is critical in determining the resection method and surveillance interval; however, the accuracy and reliability of existing measurement methods are suboptimal. We establish a computer-aided measuring (CAM) system to automatically and accurately measure colorectal polyp size in real time.

We conducted a hybrid retrospective and prospective preclinical study to develop models retrospectively and validate them prospectively. The CAM system includes two deep learning models: a polyp detection model trained on 15,620 images and tested on 4,438 images and a depth-map prediction model trained on 18,500 red-green-blue depth-map pairs and tested on 1,500 pairs. A digitized three-dimensional model development setting (3D-MDS) was used to validate the CAM system; here, 3,750 red-green-blue images from 15 digital virtual 3D polyps were generated for size measurements. A high-simulation colon/polyp model was used to test real-time polyp size measurement performance.

In the 3D-MDS, the CAM system had an average relative error of 7.9% (standard deviation 4.19%) for the depth-map prediction model and a mean percentage error of 7.9% (interquartile range [IQR] 4.6%-10.4%) for the polyp size measurement. In the high-simulation colon/polyp model, the CAM system's depth distance measurement had a percentage error of 12.0% (IQR 5.5%-15.7%). For the polyp size measurement, the highest accuracy and reliability were achieved at the medium-depth distance and the position within the center field-of-view, with percentage errors of 4.1% (IQR 2.6%-4.5%) and 3.3% (IQR 1.8%-3.5%) for the first and second measurements, respectively, and an intraclass correlation coefficient of 0.999.

The CAM system showed excellent accuracy and reliability in diverse testing scenarios; here, the medium-depth distance (15-35 mm) and the position within the center field-of-view was likely the optimal measurement location.