Conditional Score-Based Diffusion Model for Cortical Thickness Trajectory Prediction

TL;DR

This work proposes a conditional score-based diffusion model specifically designed to generate CTh trajectories with the given baseline information, such as age, sex, and initial diagnosis, and demonstrates an uncertainty analysis of patient-specific CTh prediction through multiple realizations.

Abstract

Alzheimer's Disease (AD) is a neurodegenerative condition characterized by diverse progression rates among individuals, with changes in cortical thickness (CTh) closely linked to its progression. Accurately forecasting CTh trajectories can significantly enhance early diagnosis and intervention strategies, providing timely care. However, the longitudinal data essential for these studies often suffer from temporal sparsity and incompleteness, presenting substantial challenges in modeling the disease's progression accurately. Existing methods are limited, focusing primarily on datasets without missing entries or requiring predefined assumptions about CTh progression. To overcome these obstacles, we propose a conditional score-based diffusion model specifically designed to generate CTh trajectories with the given baseline information, such as age, sex, and initial diagnosis. Our conditional diffusion model utilizes all available data during the training phase to make predictions based solely on baseline information during inference without needing prior history about CTh progression. The prediction accuracy of the proposed CTh prediction pipeline using a conditional score-based model was compared for sub-groups consisting of cognitively normal, mild cognitive impairment, and AD subjects. The Bland-Altman analysis shows our diffusion-based prediction model has a near-zero bias with narrow 95% confidential interval compared to the ground-truth CTh in 6-36 months. In addition, our conditional diffusion model has a stochastic generative nature, therefore, we demonstrated an uncertainty analysis of patient-specific CTh prediction through multiple realizations.

Figures (3)

• Figure 1: Overview of the score-based diffusion model framework for cortical thickness (CTh) prediction. (A) Training of the conditional score-based diffusion model is achieved through a forward diffusion process integrating baseline information such as demographics, diagnosis, and inter-scan interval with CTh measurements. The baseline and follow-up T1-weighted MRIs are pre-processed to extract the CTh using FreeSurfer. The concatenated baseline information is then fed through the 1D denoising U-net to estimate the score function, which is used in the reverse diffusion process. The 1D denoising U-net is trained by MSE loss to estimate gradual changes in data distributions from baseline information to prediction of CTh changes. (B) The prediction stage employs a reverse diffusion process conditioned on baseline characteristics and CTh, utilizing a trained denoising U-net and iterative denoising to estimate changes in CTh, which can then be used to predict future CTh trajectories. Note that our framework supports continuous prediction through a flexible selection of time differences.
• Figure 2: Bland-Altman analysis and correlation of predicted and measured cortical thickness across different subgroups. (A) From left to right, Bland-Altman plots show the agreement between predicted and actual cortical thickness for all subjects (N=178), CN subjects (N=70), MCI subjects (N=68), and AD subjects (N=40), respectively, with mean differences (MD) near zero predictive bias. (B) Scatter plots with linear regression analysis demonstrate strong correlations (R$^2>$ 0.9) between predicted and actual cortical thickness across all subjects and all subgroups.
• Figure 3: Longitudinal cortical thickness predictions in AD and CN exemplar subjects. (A) Continuous prediction versus actual sparse measurement of average cortical thickness in a 60-year-old male AD patient and 71-year-old male CN subject over 36 months. Predicted values for AD (red) demonstrate more obvious cortical thinning, while data for CN subjects (black) exhibit relative stability. (B) Separate analyses of the left and right para-hippocampal regions show a similar pattern, with the AD patient exhibiting a more pronounced decline in cortical thickness than the CN subject. The error bar of predicted values demonstrates the uncertainty of the model through multiple realizations.