Prompt-Driven Latent Domain Generalization for Medical Image Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Deep learning models in medical imaging are prone to distribution shifts due to artifacts and variations, compromising diagnostic accuracy.
• Existing domain generalization (DG) methods often require accurate domain labels, which are not always available for medical datasets.
• This limitation hinders the development of robust AI models for real-world clinical applications.

Purpose of the Study:

• To propose a unified domain generalization framework (PLDG) for medical image classification that operates without requiring domain labels.
• To enhance the reliability and generalizability of deep learning models in medical image analysis.
• To address the challenge of distribution shifts in medical imaging datasets.

Main Methods:

• Developed Prompt-driven Latent Domain Generalization (PLDG), a framework combining unsupervised domain discovery and prompt learning.
• Implemented pseudo domain label generation via clustering of style features.
• Utilized collaborative domain prompts and a Vision Transformer for cross-domain knowledge learning, enhanced by a domain prompt generator and domain mixup strategy.

Main Results:

• PLDG achieved comparable or superior performance to conventional DG algorithms across three medical image classification tasks and one debiasing task.
• The framework successfully generalized to unseen domains without relying on explicit domain labels.
• Demonstrated the effectiveness of unsupervised domain discovery and prompt learning for robust medical image analysis.

Conclusions:

• Prompt-driven Latent Domain Generalization (PLDG) offers an effective solution for training reliable deep learning models in medical imaging without domain labels.
• The proposed method addresses key limitations of existing DG techniques, paving the way for more robust AI diagnostics.
• PLDG shows significant potential for improving the clinical applicability of AI in healthcare.