Measurement-Conditioned Denoising Diffusion Probabilistic Model for Under-Sampled Medical Image Reconstruction

Yutong Xie, Yutong Xie 0004, Quanzheng Li

TL;DR

This paper introduces Measurement-Conditioned Denoising Diffusion Probabilistic Models (MC-DDPMs) for reconstructing medical images from undersampled data. By conditioning the diffusion process directly on the undersampled measurements and the sampling mask, MC-DDPMs achieve state-of-the-art reconstruction quality and can also quantify reconstruction uncertainty.

What Problem Does This Paper Address?

Medical imaging techniques like MRI often require acquiring data in a sparse or undersampled manner to reduce scan times, patient discomfort, or radiation exposure. However, this undersampling leads to artifacts and a loss of image quality. Reconstructing a high-quality image from this limited data is a significant challenge. Traditional methods often struggle to balance fidelity to the acquired measurements with the generation of realistic image content. While recent advances in diffusion models have shown promise for image generation and reconstruction, they typically operate in the image domain and don't inherently incorporate the undersampling information directly into the diffusion process. This paper aims to bridge this gap by developing a diffusion model that is explicitly conditioned on the undersampled measurements and the undersampling mask, enabling more accurate and robust reconstruction.

Methods

The core innovation is the Measurement-Conditioned Denoising Diffusion Probabilistic Model (MC-DDPM). Unlike standard DDPMs that operate on the image domain, MC-DDPM defines the diffusion and reverse processes in the measurement domain (e.g., k-space for MRI). This means that instead of adding noise to an image and then learning to denoise it, the model adds noise to the undersampled measurements and learns to reverse this process.

Key aspects of the method include:

  • Measurement Domain Diffusion: The forward diffusion process adds Gaussian noise to the undersampled measurements ($y_{M^c, t}$), gradually transforming them into pure noise. The reverse process learns to denoise these measurements step-by-step.
  • Conditioning on Undersampling Mask: The diffusion process is conditioned on the undersampling mask ($M^c$), ensuring that the model respects the known sampled data and learns to infer the missing information. This conditioning is integrated directly into the model, eliminating the need for separate data consistency steps during sampling.
  • Generative Nature for Uncertainty Quantification: Because MC-DDPM is a generative model, it can produce multiple plausible reconstructions from the same undersampled measurements. The variation among these samples can be used to quantify the uncertainty associated with the reconstruction, providing valuable information about the reliability of different image regions.
  • Unified Framework: The method is presented as a unified framework that can handle various undersampled reconstruction tasks by defining the measurement domain appropriately.

Datasets

The experiments were conducted using the fastMRI single-coil knee dataset. This dataset is publicly available and was used for both training and evaluation. The authors mention that their method can be applied to other undersampled medical imaging tasks, implying potential applicability to datasets from other modalities like CT.

Key Insights and Ideas

  • Domain Matters: Performing the diffusion process in the measurement domain (k-space) rather than the image domain is crucial for effectively incorporating undersampling information and achieving better reconstruction.
  • Implicit Data Consistency: Conditioning the diffusion process on the undersampling mask naturally enforces data consistency, removing the need for explicit post-hoc correction steps that are common in other methods.
  • Uncertainty is Quantifiable: The generative nature of MC-DDPM allows for the estimation of reconstruction uncertainty by analyzing the variance across multiple generated samples. This is a significant advantage for clinical applications where knowing the reliability of a reconstruction is important.
  • Flexibility of DDPM: The DDPM framework, with its flexible noise distribution control, proves to be a powerful tool for tackling complex inverse problems like undersampled image reconstruction.

Key Results

MC-DDPM demonstrated state-of-the-art performance on accelerated MRI reconstruction tasks.

  • Quantitative Improvements: For 4x acceleration, MC-DDPM achieved a PSNR of 36.69 and SSIM of 0.905 on PD data, outperforming U-Net and zero-filled reconstruction. For 8x acceleration, it achieved a PSNR of 33.49 and SSIM of 0.862.
  • Comparison to Score-Based Models: The method outperformed a state-of-the-art score-based reconstruction method by up to 3.62 PSNR and 0.077 SSIM for 4x acceleration, and 2.96 PSNR and 0.089 SSIM for 8x acceleration.
  • Qualitative Results: Visualizations showed that MC-DDPM produced more realistic structures and fewer artifacts compared to baseline methods, especially at higher acceleration factors.
  • Uncertainty Visualization: The standard deviation maps generated by MC-DDPM highlighted regions with higher reconstruction uncertainty, which often corresponded to areas with more complex structures or finer details.

Limitations and Future Work

  • Variance Schedule: The authors note that the variance schedule ($\beta_t$) is an important hyperparameter that influences sampling quality and efficiency. Further investigation into optimal variance schedules is suggested.
  • Generalizability: While demonstrated on MRI, the authors believe MC-DDPM can be applied to other undersampled medical imaging tasks like limited-angle or sparse-view CT reconstruction, suggesting this as a future research direction.
  • Hyperparameter Tuning: The paper mentions that the diffusion process is sensitive to hyperparameters, and further exploration of their impact is warranted.

Summary updated

Abstract

We propose a novel and unified method, measurement-conditioned denoising diffusion probabilistic model (MC-DDPM), for under-sampled medical image reconstruction based on DDPM. Different from previous works, MC-DDPM is defined in measurement domain (e.g. k-space in MRI reconstruction) and conditioned on under-sampling mask. We apply this method to accelerate MRI reconstruction and the experimental results show excellent performance, outperforming full supervision baseline and the state-of-the-art score-based reconstruction method. Due to its generative nature, MC-DDPM can also quantify the uncertainty of reconstruction. Our code is available on github.
BibTeX
@inproceedings{xie2022measurement,
  author = {Yutong Xie and Yutong Xie and Quanzheng Li},
  title = {{Measurement-Conditioned Denoising Diffusion Probabilistic Model for Under-Sampled Medical Image Reconstruction}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  doi = {10.1007/978-3-031-16446-0\_62},
}
Figure