Steerable Conditional Diffusion for Out-of-Distribution Adaptation in Medical Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Denoising diffusion models are widely used for inverse problems in medical imaging.
• Their performance on out-of-distribution (OOD) tasks, where data differs from training, is a significant challenge.
• OOD data can lead to reconstructions with hallucinated features specific to the training set.

Purpose of the Study:

• To address the challenge of hallucinated features in diffusion model reconstructions for OOD imaging tasks.
• To improve the accuracy and robustness of diffusion models when applied to data not seen during training.
• To introduce a novel test-time adaptation framework for diffusion models in imaging.

Main Methods:

• Introduced Steerable Conditional Diffusion, a test-time adaptation sampling framework.
• The framework adapts the diffusion model concurrently with image reconstruction.
• Adaptation is guided solely by information from the available measurement.

Main Results:

• Achieved substantial enhancements in OOD performance across diverse imaging modalities.
• Demonstrated significant improvements in reconstruction accuracy for OOD datasets.
• Successfully reduced the hallucination of training-specific image features in reconstructions.

Conclusions:

• Steerable Conditional Diffusion effectively improves the OOD performance of denoising diffusion models.
• The proposed method enhances reconstruction accuracy by adapting models at test time.
• This work advances the reliable application of diffusion models in real-world imaging scenarios.