Automated detection of cecal intubation with variable bowel preparation using a deep convolutional neural network

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Gastroenterology Technology

Background:

• Colonoscopy completion is crucial for reducing post-colonoscopy colorectal cancer rates.
• Automated appendiceal orifice (AO) detection using artificial intelligence (AI) is explored for quality assurance.
• Existing AI algorithms' performance in suboptimal conditions like variable bowel preparation remains undemonstrated.

Purpose of the Study:

• To develop and evaluate an automated computer-assisted method for detecting the AO.
• To ensure AO detection irrespective of varying bowel preparation quality.
• To utilize a deep convolutional neural network (CNN) for robust AO identification.

Main Methods:

• Extracted 13,222 images from 35 colonoscopy videos (2015-2018), labeled with Boston Bowel Preparation Scale (BBPS) scores.
• Trained a CNN using a DenseNet architecture on 11,900 images and tested on 1,322 images.
• Employed binary cross-entropy loss for training a classifier to distinguish AO from non-AO images.

Main Results:

• The deep CNN achieved 94% accuracy in classifying AO and non-AO images, with an AUC of 0.98.
• Algorithm demonstrated high performance: 96% sensitivity, 92% specificity, 92% PPV, and 96% NPV.
• AO detection exceeded 95% across all BBPS scores; non-AO detection improved from 83.95% (BBPS 1) to 98.28% (BBPS 3).

Conclusions:

• A deep CNN effectively discriminates AO from non-AO images, even with suboptimal bowel preparation.
• The developed algorithm shows significant potential for enhancing colonoscopy quality assurance.
• Further real-time testing is required to validate the algorithm's effectiveness in clinical practice.