Regression Is All You Need for Medical Image Translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Generative Adversarial Nets (GANs) and Diffusion Models (DMs) excel at natural image synthesis but can introduce inaccuracies like hallucinations and noise in medical imaging.
• Accuracy and fidelity are critical in medical applications, making standard generative models potentially unsuitable.

Purpose of the Study:

• To introduce YODA (You Only Denoise once - or Average), a 2.5D diffusion-based framework for medical image translation (MIT).
• To address the noise replication issue in conventional diffusion sampling by proposing Expectation-Approximation (ExpA) sampling and a novel regression sampling method.

Main Methods:

• YODA utilizes a diffusion-based framework with a novel regression sampling technique that omits iterative refinement for single-step, noise-free image generation.
• Expectation-Approximation (ExpA) sampling, involving averaging multiple samples, was explored to mitigate noise inherent in standard diffusion sampling.

Main Results:

• Regression sampling in YODA demonstrated substantial efficiency gains and matched or surpassed the image quality of full diffusion sampling, even with ExpA.
• Iterative refinement was found to enhance perceptual realism but not information translation, as confirmed by downstream task performance.
• YODA outperformed eight state-of-the-art DMs and GANs, challenging the notion that complex models are always superior for high-quality MIT.

Conclusions:

• YODA's regression sampling offers an efficient and effective approach to medical image translation, producing high-fidelity images.
• The framework challenges the dominance of complex generative models, showing comparable or superior performance to computationally cheaper regression models.
• YODA-translated images are suitable for medical applications, being interchangeable with or even superior to physical acquisitions.