Automated BBPS Scoring in Colonoscopy: A Comparative Analysis of Pre-trained Deep Learning Architectures

Presently, an effective computer-aided system to assess the quality of bowel preparation in colonoscopy is lacking. The present work focuses on the development of automated Boston Bowel Preparation Scale (BBPS) segmental score-based classification for an accurate assessment of the quality of bowel preparation in colonoscopy. Five different deep learning architectures, namely ResNet-50, ResNet-101, MobileNet-V2, Xception, and Inception-V3, were applied to benchmark the existing dataset in this research area. A comparative analysis has been shown using various evaluation metrics, test set analysis, and interpretability plots. The results indicate that ResNet-50 achieved the highest performance, with 100% AUC, 100% F1 score, and 100% accuracy on the training data, and 96.67% AUC, 93.89% F1 score, and 93.89% accuracy on the validation data. ResNet-101, MobileNet-V2, and Xception also delivered robust results, while Inception-V3 lagged behind. The findings demonstrate the effectiveness of deep learning architectures in the automatic assessment of bowel cleanliness in colonoscopy procedures.