Endoscopic diagnosis and treatment planning for colorectal polyps using a deep-learning model

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate colorectal polyp histology prediction is crucial for patient management.
• Distinguishing between serrated polyps, benign adenomas, and submucosal cancers requires expertise.
• Current diagnostic methods can be subjective and vary in accuracy.

Purpose of the Study:

• To develop and validate a deep-learning computer-aided diagnostic (CAD) system for predicting colorectal polyp histology.
• To compare the CAD system's performance against endoscopist expertise levels.

Main Methods:

• A deep-learning model was trained on 12,480 narrow-band imaging (NBI) image patches from 624 colorectal polyps.
• The CAD system was validated on two independent datasets comprising 545 polyps.
• Performance was evaluated by comparing CAD predictions to true histology (serrated polyp, benign adenoma/mucosal or superficial submucosal cancer, deep submucosal cancer).

Main Results:

• The CAD system achieved an overall kappa value of 0.614-0.642, comparable to experts (0.649-0.735) and superior to trainees (0.368-0.401).
• Areas under the receiver operating curves ranged from 0.86-0.95 across polyp categories.
• Overall diagnostic accuracy was 81.3-82.4%, significantly outperforming trainees (63.8-71.8%) and matching expert levels.

Conclusions:

• Deep-learning-based CAD systems can accurately assess colorectal polyp histology.
• The CAD system demonstrated performance comparable to experienced endoscopists.
• CAD assistance significantly improved diagnostic accuracy and kappa values for less experienced endoscopists, potentially enhancing clinical decision-making.