Functional Imaging Constrained Diffusion for Brain PET Synthesis from Structural MRI

TL;DR

Quantitative and qualitative analyses conducted on 293 subjects with paired T1-weighted MRI and 18F-fluorodeoxyglucose-PET scans suggest that FICD achieves superior performance in generating FDG-PET data compared to state-of-the-art methods.

Abstract

Magnetic resonance imaging (MRI) and positron emission tomography (PET) are increasingly used in multimodal analysis of neurodegenerative disorders. While MRI is broadly utilized in clinical settings, PET is less accessible. Many studies have attempted to use deep generative models to synthesize PET from MRI scans. However, they often suffer from unstable training and inadequately preserve brain functional information conveyed by PET. To this end, we propose a functional imaging constrained diffusion (FICD) framework for 3D brain PET image synthesis with paired structural MRI as input condition, through a new constrained diffusion model (CDM). The FICD introduces noise to PET and then progressively removes it with CDM, ensuring high output fidelity throughout a stable training phase. The CDM learns to predict denoised PET with a functional imaging constraint introduced to ensure voxel-wise alignment between each denoised PET and its ground truth. Quantitative and qualitative analyses conducted on 293 subjects with paired T1-weighted MRI and 18F-fluorodeoxyglucose (FDG)-PET scans suggest that FICD achieves superior performance in generating FDG-PET data compared to state-of-the-art methods. We further validate the effectiveness of the proposed FICD on data from a total of 1,262 subjects through three downstream tasks, with experimental results suggesting its utility and generalizability.

Figures (8)

• Figure 1: Illustration of the proposed functional imaging constrained diffusion (FICD) framework for 3D brain PET image synthesis with paired structural MRI as the condition. (a) The training phase consists of a forward diffusion process that incrementally adds noise to an input PET image, and a generative reverse denoising process that gradually removes the noise. The MRI condition and noise-corrupted PET are input to the proposed constrained diffusion model (CDM) to predict the noise on PET, then the noise is further used to estimate the denoised PET in the previous timestep and in the final timestep. The outputs of CDM are optimized by a unique functional imaging constraint and a noise-level constraint, respectively. (b) In the inference phase, an MRI and a pure noise are input to the CDM, which progressively removes the noise to generate a synthetic PET image.
• Figure 2: Synthetic PET images of the same subject (ID: 002_S_4270) from ADNI jack2008alzheimer generated through eight separate sampling iterations by (a) DDPM ho2020denoising and (b) our FICD, with each iteration introducing new random noise. The input MRI and the ground-truth PET are displayed on the left.
• Figure 3: Visualization of PET images synthesized by the proposed FICD and six competing methods on cognitively normal subjects from the test set in ADNI jack2008alzheimer. The ground-truth PET images and input MRI are displayed at the bottom with the corresponding subject ID.
• Figure 4: Difference maps for synthetic PET images generated by the proposed FICD and six competing methods on cognitively normal subjects from the test set in ADNI jack2008alzheimer. The ground-truth PET images and input MRI are displayed at the bottom with the corresponding subject ID.
• Figure 5: Scatter diagrams of cognitive scores ( i.e., MMSE and CDR) predicted by different methods for CLAS-SCD at 7-year follow-up and ADNI-SMC at 2-year follow-up, with baseline images as input. CC: correlation coefficient; MMSE: mini-mental state examination; CDR: clinical dementia rating.