Polyp size prediction with an artificial intelligence based monocular depth estimation method

Aims Accurate measurement of the polyp size in colonoscopy is important for choosing the guideline recommended resection method and surveillance interval. However, size estimations of endoscopists can be off by up to 65%. We aimed to develop an artificial intelligence (AI) based Method to improve size estimation.

Methods We used a real scale 3D colon model with simulated 3D polyps to render 3060 white light images of a virtual colonoscopy. For every image, we also rendered the corresponding depth map. The virtual white light images and depth maps were then used to finetune a pretrained monocular depth estimation model. For testing or model, we used single images of 133 lesions, which were treated by endoscopic mucosa dissection (ESD) in our department. For comparison, three ESD trained expert endoscopists also estimated the lesion size.

Results The mean absolute percentage error (MAPE) for our AI size estimation model was 29%. The average of the three endoscopists’ MAPE was 28%.

Conclusions Monocular depth estimation is a promising AI method for automatic poly size prediction with comparable accuracy to expert endoscopists

Publication History

Article published online:
27 March 2025