Wireless capsule endoscopy multiclass classification using three-dimensional deep convolutional neural network model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) offers non-invasive gastrointestinal tract visualization.
• Manual review of WCE frames is time-consuming and error-prone.
• Existing computer-aided diagnosis (CAD) systems often neglect temporal information in WCE videos.

Purpose of the Study:

• To develop an automatic multiclass classification system for WCE diagnosis.
• To leverage spatiotemporal information for improved WCE abnormality detection.
• To enhance the efficiency and accuracy of WCE analysis.

Main Methods:

• A three-dimensional deep convolutional neural network (3D-CNN) was proposed.
• The 3D-CNN model processed sequential WCE frames to capture spatiotemporal data.
• Performance was evaluated against 2D-CNN and pre-trained networks using 29 WCE videos (14,691 frames).

Main Results:

• The 3D-CNN model achieved superior performance across all metrics.
• Sensitivity: 98.92% (3D-CNN) vs. 98.05% (2D-CNN).
• Specificity: 99.50% (3D-CNN) vs. 86.94% (2D-CNN).
• Accuracy: 99.20% (3D-CNN) vs. 92.60% (2D-CNN).

Conclusions:

• The proposed 3D-CNN model significantly outperforms 2D-CNN and pre-trained networks.
• The model effectively utilizes temporal and spatial information for accurate WCE lesion detection.
• This approach offers an efficient tool for improving WCE diagnosis.