STAR-ViT: A Spatially Transformed and Adversarially Realigned Vision Transformer for Pathogen Classification in...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Accurate computer-aided diagnosis of pediatric pneumonia is hindered by insufficient annotated data.
• Existing methods struggle with cross-modal generalization between different imaging types like CT and X-ray.

Purpose of the Study:

• To develop an advanced vision transformer model, STAR-ViT, for improved pediatric pneumonia diagnosis.
• To enhance robustness and cross-modal generalization using novel architectural components.

Main Methods:

• Proposed STAR-ViT integrates spatial transformation consistency and adversarial cross-modal realignment.
• A task-coordinated modulation (TCM) module stabilizes multi-objective encoder optimization.
• Causal feature modeling separates disease features from confounding variations.

Main Results:

• STAR-ViT achieved 92.07% accuracy on a private CT dataset for pediatric pneumonia pathogen classification.
• Achieved 97.86% accuracy on the public ChestXRay2017 dataset for pneumonia classification.
• Outperformed existing state-of-the-art methods in experimental evaluations.

Conclusions:

• STAR-ViT demonstrates superior performance in pediatric pneumonia diagnosis.
• The proposed methods effectively address data limitations and improve cross-modal generalization.
• This AI model offers a promising tool for enhancing medical image analysis in pediatric care.