用于神经影像规范建模的去噪扩散网络

Normative modeling is a crucial tool in neuroimaging for identifying individual deviations from a healthy population, yet existing methods often struggle with complex, high-dimensional data, particularly in capturing non-linear relationships and multimodal distributions. This paper introduces a novel normative modeling framework based on denoising diffusion probabilistic models (DDPMs) to address these limitations. By learning the intricate structure of data distributions, this framework can generate high-quality synthetic data and accurately estimate expected values of brain structural or functional metrics, conditioned on covariates like age and sex. Compared to traditional methods such as Gaussian Process Regression, the DDPM model demonstrates significant advantages in capturing non-Gaussian and multimodal characteristics of data distributions, thereby enhancing the sensitivity and specificity of anomaly detection. Experimental results show superior performance across multiple neuroimaging datasets, offering a new perspective for early diagnosis and personalized treatment of neuropsychiatric disorders. Furthermore, the generative capabilities of DDPMs open new avenues for data augmentation and privacy-preserving research, promising to advance the widespread application of normative modeling in both clinical and research settings. This approach provides a robust and flexible solution for understanding individual brain differences and their implications for health and disease, moving beyond the constraints of simpler statistical models.