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

Area of Science:

• Veterinary Medicine
• Oncology
• Biomedical Engineering

Background:

• Small animal models, especially mice, are vital for gastrointestinal disease research, including colorectal cancer.
• Colonoscopy in these models generates extensive video data, making manual analysis resource-intensive and time-consuming.
• Automated analysis is essential for efficient tumor assessment in preclinical endoscopic studies.

Purpose of the Study:

• To develop and evaluate an automated deep learning system for detecting and classifying colon tumors in mouse colonoscopy videos.
• To improve the efficiency and accuracy of tumor assessment in preclinical research.
• To provide real-time analytical support for researchers conducting endoscopic studies in mouse models.

Main Methods:

• A YOLOv7 deep learning model, pre-trained on human polyp images, was adapted for mouse colonoscopy video analysis.
• Tumor detection was optimized using a complementary stool detector and a color-based filter.
• Lesion classification employed a custom ratio-based method, categorizing tumors from '0' (none) to '5' (>50% colon diameter).

Main Results:

• The YOLOv7 model achieved initial detection Precision of 0.576 and Recall of 0.916.
• Integration of the stool detector significantly improved detection metrics to Precision 0.932, Recall 0.946, and Accuracy 0.897.
• Tumor classification accuracy against expert annotations reached 0.759 Precision, 0.774 Recall, and 0.774 Accuracy across all five classes.

Conclusions:

• The developed deep learning system demonstrates reliable performance in detecting and classifying colon tumors in mice.
• This automated approach offers valuable real-time support for preclinical endoscopic research.
• Further validation studies are recommended to fully ascertain the system's capabilities and potential clinical impact.