YOLO-Based Deep Learning Model for Real-Time Polyp and Dysplasia Detection in IBD Colonoscopy

Introduction
Colorectal cancer is the second leading cause of cancer deaths worldwide. Patients with inflammatory bowel disease (IBD) face a heightened risk of dysplasia and colorectal cancer. Dysplastic lesions in IBD are often flat, ill-defined and located on inflamed or scarred mucosa, making detection and optical diagnosis highly operator-dependent and limiting the effectiveness of current surveillance programs. This study presents a deep learning tool for automatic detection of polyps and dysplastic lesions in colonoscopy images from IBD patients.

Methods
A diverse training dataset was constructed from ten public colonoscopy repositories, totalling 19,443 images. Additionally, 858 expert-annotated IBD-specific images were provided by Hospital Universitari i Politècnic La Fe of Valencia. A YOLOv11s-based model was developed using transfer learning, optimised through hyperparameter tuning with Optuna, and refined via fine-tuning on IBD-specific data. Performance was evaluated using 5-fold cross-validation and benchmarked against three IBD-focused models: RetinaNet, EfficientDet, and CSPDarkNet.

Results
The final model was evaluated using cross-validation, with all evaluation images drawn exclusively from the IBD dataset. It achieved a precision of 71.7% and a recall of 64.1%, representing improvements of 17% and 31%, respectively, over a baseline model trained only on public colonoscopy datasets. Recall for dysplastic lesions reached 84.3%. Evaluated on colonoscopy videos, the model demonstrated robust and consistent real-time detection, maintaining performance under challenging conditions such as partial visibility, tissue folds, and colour variation. Compared to benchmark models, YOLO showed competitive metrics with superior real-time processing capability.

Conclusions
The proposed YOLO-based model shows promising performance for automated IBD lesion detection, with particular strength in identifying clinically critical dysplastic lesions. Its real-time capability and low computational requirements make it suitable for clinical integration as a decision-support tool during colonoscopy.