Automated lesion detection in endoscopic imagery for small animal models - a pilot study.

Abstract

Objectives: Small animal models, particularly mice, are crucial for studying gastrointestinal diseases like colorectal cancer. Tumor assessment via colonoscopy generates large video datasets, necessitating automated analysis due to limited resources and time-consuming manual review.

Methods: We employed a YOLOv7-based deep learning model pre-trained on human polyp images to detect tumors in mouse colonoscopy videos. Detection was enhanced using a stool detector and a color-based filter. Lesions were classified from '0' (no tumor) to '5' (tumor >50 % of colon diameter) using a custom ratio-based method. The system was evaluated on 150 videos from 28 mice over 6 weeks, with 125 videos containing tumors.

Results: Initial detection yielded a Precision of 0.576, Recall of 0.916, and Accuracy of 0.593. Adding the stool detector improved results to 0.932, 0.946, and 0.897, respectively. Compared to expert annotations, classification reached 0.759 Precision, 0.774 Recall, and 0.774 Accuracy over all five classes.

Conclusions: The proposed approach reliably detects and classifies colon tumors in mice, offering real-time support for preclinical endoscopic studies. Further evaluation will provide more insights into its performance.