[Paper Review] Brain PET Synthesis from MRI Using Joint Probability Distribution of Diffusion Model at Ultrahigh Fields

Xie Taofeng, Cao Chentao+9 more|arXiv (Cornell University)|Nov 16, 2022

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper proposes a joint probability diffusion model (JPDDM) to synthesize brain PET scans from MRI at ultrahigh fields (5T and 7T), leveraging conditional score estimation via a UNet to model the joint distribution of MRI and PET. The method achieves superior PSNR over CycleGAN and score-based SDE baselines, demonstrating high-fidelity PET synthesis with improved stability and accuracy.

ABSTRACT

MRI and PET are important modalities and can provide complementary information for the diagnosis of brain diseases because MRI can provide structural information of brain and PET can obtain functional information of brain. However, PET is usually missing. Especially, simultaneous PET and MRI imaging at ultrahigh field is not achievable in the current. Thus, synthetic PET using MRI at ultrahigh field is essential. In this paper, we synthetic PET using MRI as a guide by joint probability distribution of diffusion model (JPDDM). Meanwhile, We utilized our model in Ultrahigh Fields.

Motivation & Objective

• Address the clinical challenge of limited PET availability due to cost and radiation exposure by synthesizing PET from MRI.
• Enable cross-modal synthesis of PET at ultrahigh-field MRI (5T and 7T), where simultaneous PET-MRI is currently infeasible.
• Improve generation stability and accuracy in MRI-to-PET synthesis by modeling the joint probability distribution of MRI and PET via diffusion models.
• Overcome limitations of existing methods like CycleGAN and score-based SDEs by incorporating conditional guidance and joint distribution learning.

Proposed method

• Formulate the PET synthesis task as learning the conditional score function ∇ₓₚₑₜ log p(xₚₑₜ|xₘᵣᵢ) via joint diffusion modeling of MRI and PET.
• Implement a forward SDE that gradually corrupts the joint MRI-PET data into noise, with noise scales σₘᵢₙ to σₘₐₓ.
• Use a UNet-based score network sθ*(Xᵢ₊₁, i+1) to estimate the score ∇ₓ log pₜ(Xᵢ), trained via a denoising objective L(θ;σ) = ½𝔼[‖σₘᵢₙ(σₘₐₓ/σₘᵢₙ)ᵗ sθ(Xᵢ₊₁,σ) + z‖₂²].
• Apply a predictor-corrector sampling scheme: predictor uses reverse SDE with drift fᵢ = 0 and diffusion gᵢ = √(σ²ᵢ − σ²ᵢ₋₁), while corrector applies Langevin dynamics with step size ε.
• Condition the generation on MRI by fixing xₘᵣᵢ during sampling, ensuring the model synthesizes PET consistent with input MRI.
• Train and evaluate on the ADNI dataset (14,440 registered MRI-PET pairs), with images resized to 128×128.

Experimental results

Research questions

• RQ1 Can a joint diffusion model effectively synthesize high-fidelity brain PET images from MRI at ultrahigh fields (5T and 7T)?
• RQ2 How does modeling the joint probability distribution of MRI and PET improve generation quality and stability compared to conditional GANs or score-based SDEs?
• RQ3 What is the performance of the proposed JPDDM in terms of PSNR and visual quality when applied to real 5T and 7T MRI scans?
• RQ4 Does the method maintain robustness and accuracy when generalizing to high-resolution, high-SNR ultrahigh-field MRI data?

Key findings

• The proposed JPDDM achieved higher Peak Signal-to-Noise Ratio (PSNR) than both CycleGAN and score-based SDE baselines on the ADNI dataset.
• Visual comparisons in Figure 2 show that the JPDDM produces more realistic and detailed synthetic PET images with fewer artifacts.
• The method successfully synthesized PET images from 5T and 7T MRI scans, as demonstrated in Figures 3 and 4, confirming its applicability to ultrahigh-field data.
• The joint modeling of MRI and PET via diffusion improved generation stability and fidelity, particularly in preserving metabolic patterns in synthetic PET.
• Despite superior image quality, the method exhibits slower inference speed due to iterative sampling in the predictor-corrector process.
• The model’s performance was validated on real clinical 5T and 7T MRI scanners, confirming its potential for real-world neuroimaging applications.