Video Analysis of Small Bowel Capsule Endoscopy Using a Transformer Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) is effective for small bowel disease detection but faces challenges with time-consuming image analysis and variable accuracy.
• Current deep learning methods often analyze individual WCE frames, potentially missing crucial temporal information and operator-dependent frame selection biases.

Purpose of the Study:

• To develop an automated lesion detection system for WCE that utilizes the entire video, overcoming limitations of manual frame selection and individual image analysis.
• To enhance the accuracy and efficiency of diagnosing small bowel diseases through advanced deep learning techniques.

Main Methods:

• Proposed a novel Transformer-architecture-based neural encoder designed to process entire WCE videos as input.
• The system leverages the Transformer's ability to capture long-term global correlations within and between sequential frames, extracting temporal context and intra-frame attentional features.

Main Results:

• Achieved high performance on benchmark WCE datasets, demonstrating 95.1% sensitivity and 83.4% specificity in lesion detection.
• The method effectively utilizes information from all frames and their temporal relationships, surpassing the limitations of analyzing only selected images.

Conclusions:

• The proposed Transformer-based approach significantly advances automated lesion detection in WCE by analyzing whole videos.
• This method holds the potential to improve diagnostic accuracy and efficiency in identifying small bowel diseases, reducing reliance on operator expertise.