Detecting Pneumonia using Convolutions and Dynamic Capsule Routing for Chest X-ray Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Capsule Networks (CapsNet) represent a state-of-the-art approach for image classification.
• Pneumonia detection from chest X-ray (CXR) images is a critical diagnostic task.

Purpose of the Study:

• To investigate the efficacy of CapsNet for detecting pneumonia in CXR images.
• To develop and evaluate novel convolutional capsule network architectures for improved pneumonia detection.

Main Methods:

• A simple CapsNet model was initially evaluated for pneumonia detection.
• Integration of Convolutions with Capsules (ICC) and Ensemble of Convolutions with Capsules (ECC) models were developed.
• A variant, EnCC (n=3, 4, 8, 16), was optimized, specifically E4CC.

Main Results:

• Simple CapsNet achieved comparable results to existing Deep Learning models.
• ICC and ECC models demonstrated superior performance, achieving test accuracies of 95.33% and 95.90%, respectively.
• The E4CC model achieved the highest test accuracy at 96.36% on 5857 CXR images.

Conclusions:

• Capsule Network-based models, particularly those combining convolutions and capsules, are highly effective for pneumonia detection in CXR images.
• The developed E4CC model offers a promising, accurate, and efficient tool for radiological diagnosis.
• Further research into convolutional capsule architectures can significantly advance medical image analysis.