JointDiffusion: Joint representation learning for generative, predictive, and self-explainable AI in healthcare.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Joint machine learning models for synthesis and classification often face performance and stability issues.
• Deep generative diffusion models possess internal representations valuable for both generation and prediction.

Purpose of the Study:

• To develop a stable, joint end-to-end training approach for diffusion models integrating classification and generation.
• To enhance performance in both generative and predictive tasks using shared parameterization.
• To apply the joint diffusion model to medical data challenges, including semi-supervised learning and decision explanation.

Main Methods:

• Extension of vanilla diffusion models with an integrated classifier for joint training.
• Shared parameterization between generative and classification objectives.
• Evaluation on benchmarks for classification and generation quality.
• Application to medical data, focusing on semi-supervised learning and counterfactual example generation.

Main Results:

• The proposed joint diffusion model demonstrates superior performance over state-of-the-art hybrid methods in both classification and generation.
• Achieved superior performance in a semi-supervised setting with limited human annotation.
• Successfully generated counterfactual examples for decision explanation.

Conclusions:

• The joint diffusion model offers a stable and effective approach for combined synthesis and classification tasks.
• This method shows significant promise for medical data analysis, improving performance in low-resource scenarios and providing interpretable results.