WSDC-ViT: a novel transformer network for pneumonia image classification based on windows scalable attention and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate pneumonia diagnosis is critical for effective treatment, but it presents diagnostic challenges.
• Existing computer-aided diagnosis (CADx) models have limitations in feature extraction for pneumonia detection.

Purpose of the Study:

• To introduce the WSDC-ViT network for enhanced computer-aided pneumonia detection.
• To improve diagnostic accuracy and reduce the workload for radiologists.

Main Methods:

• Developed a novel WSDC-ViT network with a scalable self-attention mechanism and convolutional refinement.
• Integrated decoupled query, key, and value dimensions for reduced computational overhead.
• Employed an interactive window-based attention module for long-range dependency modeling.
• Embedded a convolution-based module with dynamic ReLU for fine-grained local feature extraction.

Main Results:

• Achieved 95.13% accuracy and 95.63% F1-score on a chest X-ray dataset.
• Attained 99.36% accuracy and 99.34% F1-score on a CT dataset.
• Demonstrated superior performance compared to existing automated pneumonia classification methods.

Conclusions:

• The WSDC-ViT network shows significant potential for clinical application in pneumonia diagnosis.
• The proposed architecture effectively enhances both global and local feature extraction for improved accuracy.
• This AI-driven approach can aid radiologists in making faster and more precise diagnoses.