Multiple abnormality classification in wireless capsule endoscopy images based on EfficientNet using attention...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Wireless capsule endoscopy (WCE) generates vast image data, making manual lesion detection challenging.
• Existing automated methods struggle with multi-lesion detection sensitivity in WCE.
• There is a need for efficient and accurate tools to aid WCE image analysis.

Purpose of the Study:

• To develop and evaluate an AI strategy for automatic detection of multiple vascular and inflammatory lesions in WCE images.
• To improve the balance between sensitivity and speed in multi-lesion detection.
• To provide an auxiliary diagnostic tool for rapid WCE reading.

Main Methods:

• Fine-tuning EfficientNet using weakly supervised learning for feature extraction from WCE images.
• Employing an attention network that combines spatial and channel features for classification.
• Comparing the proposed model's accuracy and speed against ResNet121 and InceptionNetV4.

Main Results:

• The model achieved 96.67% sensitivity for vascular lesions and 93.33% for inflammatory lesions.
• Precision rates were 92.80% for vascular and 95.73% for inflammatory lesions.
• Achieved 96.11% accuracy, outperforming InceptionNetV4, with an image prediction time of 14 ms.

Conclusions:

• The proposed AI strategy effectively detects multiple vascular and inflammatory lesions in WCE images.
• The model demonstrates a superior balance of accuracy and speed compared to existing networks.
• This approach can serve as a valuable auxiliary tool for specialists in capsule endoscopy interpretation.