Efficacy of a comprehensive binary classification model using a deep convolutional neural network for wireless...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual review of capsule endoscopy (CE) videos for small bowel disease diagnosis is time-consuming.
• Existing automated algorithms lack sufficient validation for clinical application.
• Multi-diagnosis validation for automated CE analysis remains insufficient.

Purpose of the Study:

• To develop and test a practical binary classification model for identifying clinically significant images in CE videos.
• To evaluate the model's diagnostic accuracy on internal and external datasets.
• To assess the potential of the model for clinical-ready computer-aided reading methods.

Main Methods:

• A binary classification model was trained on 240,000 capsule endoscopy images from 84 cases.
• The model was validated and internally tested on remaining images from the training set.
• External validation was performed using 256,591 unseen images from an independent hospital.

Main Results:

• The model achieved 98.067% diagnostic accuracy on the internal validation set.
• External testing on unseen data from an independent hospital yielded 85.470% accuracy.
• The Area Under the Curve (AUC) was 0.922, with increased misreadings on external data due to ambiguous substances in 'insignificant' images.

Conclusions:

• The developed CNN-based binary classification model demonstrates excellent internal performance for CE image analysis.
• The model shows promise for clinical application but requires further refinement to address limitations with external, unseen data.
• Addressing the issue of ambiguous substances in 'insignificant' images is crucial for developing clinically-ready computer-aided reading methods.