Deep Learning-Based Real-Time Organ Localization and Transit Time Estimation in Wireless Capsule Endoscopy.

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Wireless capsule endoscopy (WCE) aids GI disease diagnosis but faces challenges with visual obstructions.
• Current machine learning models often fail with limited visual data due to reliance on color information.

Purpose of the Study:

• To develop a novel deep learning model for organ classification and transit time estimation in WCE.
• To improve WCE analysis accuracy, especially when visual information is compromised.

Main Methods:

• A deep learning model integrating Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks was developed.
• The model analyzed 2,395,932 images from 126 patients, incorporating temporal information from continuous videos.
• Gaussian filter calibration was used for boundary detection, and performance was evaluated using metrics like accuracy, MCC, and G-mean.

Main Results:

• The model achieved over 95% accuracy, sensitivity, and specificity for organ classification (stomach, small intestine, colon).
• Overall accuracy and F1-score reached 97.1%, with high MCC and G-mean values indicating robustness on imbalanced datasets.
• Transit time estimation showed mean differences of 4.3 ± 9.7 min (stomach) and 24.7 ± 33.8 min (small intestine), with 95.8% of gastric predictions within 15 min of ground truth.

Conclusions:

• The combined CNN-LSTM model is accurate and clinically effective for WCE organ classification and transit time estimation.
• Integrating temporal information enhances performance in challenging WCE scenarios.
• This approach offers a valuable tool to improve diagnostic accuracy and efficiency in managing GI diseases.