Generalized Zero-Shot Chest X-Ray Diagnosis Through Trait-Guided Multi-View Semantic Embedding With Self-Training.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Zero-shot learning (ZSL) is crucial for annotation-efficient machine learning, with significant advancements in natural image analysis.
• ZSL applications in medical imaging, particularly for chest radiographs, remain underdeveloped.
• Existing ZSL methods primarily focus on natural images, leaving a gap in medical diagnostic applications.

Purpose of the Study:

• To develop a novel strategy for generalized zero-shot diagnosis of chest radiographs.
• To explore the potential of multi-view semantic embedding for ZSL in medical imaging.
• To address challenges of noisy labels and improve performance on unseen classes in medical ZSL.

Main Methods:

• A novel strategy for generalized zero-shot diagnosis of chest radiographs was designed.
• The method leverages multi-view semantic embedding for enhanced ZSL capabilities.
• A self-training phase was incorporated to handle noisy labels and improve performance on underrepresented classes.

Main Results:

• The proposed model demonstrated consistent performance across test datasets from various sources and differing quality levels.
• Rigorous experiments confirmed the model's effectiveness in generalized zero-shot chest x-ray diagnosis.
• The method outperformed several state-of-the-art techniques in generalized zero-shot chest x-ray diagnosis.

Conclusions:

• The developed strategy offers a promising solution for generalized zero-shot diagnosis of chest radiographs.
• Multi-view semantic embedding and self-training are effective components for improving medical ZSL.
• The proposed method shows superior performance and robustness for unseen classes in chest x-ray diagnosis.