Artificial-intelligence-based QUAntification of Size measuremeNts of Adenoma in Routine Endoscopy

Objective:

The primary aim of this study is to conduct a superiority analysis between POSEIDON and endoscopist optical estimation, focusing on the errors of polyp size measurement. Secondary objectives encompass subgroup analyses concerning varying polyp sizes and histology, potential impact on follow-up protocols based on polyp sizes, and determination of the percentage of cases where obtaining measurements with POSEIDON proves unfeasible.

Study Design:

Single center, endoscopist blinded study. Considering the results of the preliminary study, a total of 42 pairs of measurements suffices to extrapolate statistical significance for a power of 90% and a significance level of 0.05. Taking dropout rates, average number of polyps per patient, polyp detection rate, the number of patients expected to be enrolled for the study is 156. For obtaining the gold standard measurement, images from the examination recording where the tip of a snare is placed next to the polyp are extracted. The snare tip and polyp are manually segmented and their largest diameter in pixels is manually calculated. The diameter of the snare-tip will be measured after the examination to account for production based variations. Based on this, the gold standard polyp size is calculated.

AI setup and limitations:

The version of the polyp detection system, EndoMind, is an updated version of the one published in (PMID: 35543874). Notably, the POSEIDON method has undergone dual refinements from its first published version (PMID: 37080235). Firstly, the AI model for water-jet tip detection has been further trained using images from colonoscopy videos, leading to performance improvements. Secondly, the pixel-to-millimeter conversion methodology has been optimized to align with the specifications of the endoscopes employed in the trial.

To obtain a polyp size estimation, initially the polyp is detected by means of a bounding box through the EndoMind system. Then, once water-jet is sprayed next to the polyp, POSEIDON detects the tip of the water-jet, where it touches the mucosa. The coordinates of the tip in the endoscopic image are correlated to the distance of the endoscope to the mucosa. With the above, the largest diameter of the bounding box is converted from pixels to millimeters, resulting in the size estimation.

For measurement acquisition, the system is initiated at the beginning of the examination and deactivated at its conclusion, facilitated by a dedicated button press. Concerning expert-system interaction, the CADe system's identifications consistently appear on an auxiliary screen within the endoscopist's workspace. On this screen, precise visualization of the impact point between the water-jet beam and the mucosa follows water-jet detection. Subsequently, the endoscopist receives visual affirmation of a successful calculation of polyp size estimation, derived from the bounding box on the screen. The endoscopist evaluates the alignment of this box around the polyp to assess the accuracy measurement outcome. Importantly, throughout these operations, the endoscopist remains blinded to the size estimated by POSEIDON. The endoscopist also reports polyp cases where they are unable to obtain a proper measurement.

Study population:

All adult patients with screening or surveillance colonoscopies that do not match the exclusion criteria will be asked for informed consent.

Intervention:

Upon polyp identification by the endoscopist, the visual assessment will be conducted, followed by obtaining a satisfactory measurement using POSEIDON when possible, as described in the AI methods section. To eliminate potential learning bias and optimize efficiency, the examiner will be blinded to the immediate measurement result. Finally, after both measurements are obtained, an instrument tip of known size will be positioned adjacent to the polyp, which is to be later used for obtaining the gold-standard polyp size.