TransMVAN: Multi-view Aggregation Network with Transformer for Pneumonia Diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate pneumonia classification from chest X-rays is vital for computer-aided diagnosis systems.
• Challenges exist in learning complex lung abnormality structures from radiographic images.

Purpose of the Study:

• To propose a novel multi-view aggregation network with Transformer (TransMVAN) for enhanced pneumonia classification in chest X-ray images.
• To improve the feature representation of lung abnormalities by incorporating glance and focus views.
• To capture complex inter-regional relationships in the lungs using a bi-directional multi-scale vision Transformer (biMSVT).

Main Methods:

• Developed TransMVAN, integrating multi-view learning (glance and focus) for enriched feature representation.
• Introduced biMSVT to model long-range dependencies and relationships between lung regions bidirectionally.
• Implemented a gated multi-view aggregation (GMVA) for adaptive feature selection from different views.

Main Results:

• Achieved Area Under the Curve (AUC) of 0.9645 and 0.9550 on two distinct chest X-ray datasets for pneumonia classification.
• Attained an AUC of 0.9761 for differentiating positive and negative polymerase chain reaction (PCR) results.
• Reached an AUC of 0.9741 for classifying normal cases, non-COVID-19 pneumonia, and COVID-19 pneumonia.

Conclusions:

• The proposed TransMVAN method demonstrates significant effectiveness in pneumonia diagnosis from chest X-ray images.
• The integration of multi-view information and advanced Transformer architecture enhances diagnostic performance.
• Experimental results validate the superiority of TransMVAN over existing comparative methods.