Voxel-level normative modeling of neuroimaging data using generative diffusion transformer
Poster No:
1518
Submission Type:
Abstract Submission
Authors:
Chang Yang1, Menghan Qin2, Weikang Gong2
Institutions:
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, 2School of Data Science, Fudan University, Shanghai, China
First Author:
Chang Yang
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University
Shanghai, China
Co-Author(s):
Introduction:
Normative modeling is an innovative statistical method that quantifies how an individual's neuroimaging measures deviate from a normative population's (Rutherford, 2022). However, existing methods are often limited to global neuroimaging measures or brain region level measures (Marquand, 2016). This may prevent from identifying potential mental disorder biomarkers and restricting the ability to conduct more detailed analysis of brain abnormalities. Recently, diffusion models have been particularly successful as generative models, especially in imaging generation task (Dhariwal,2021). In this study, we apply the latest diffusion models Diffusion Transformers (DiT, Peebles, 2023), for voxel-level normative modeling of neuroimaging data. The DiT model conditions on covariables such as Age, Sex, Brain volume (Bv), Fluid intelligence scores (Gf), years of Education (Edu), and BMI. It employs a forward diffusion and reverse denoising process to learn the distribution of voxel-level measures of healthy populations. In our experiments, we used 33,323 subjects' Fractional anisotropy (FA) images derived from Diffusion-Weighted Imaging from the UK Biobank to train the diffusion model.
Methods:
The methodology involves the following steps:
Data preparation: we collect the healthy population data with complete labels and scans. In this study, we use the FSL's FMRIB58_FA template to extract voxel data from registered scans and organize them into a one-dimensional (1D) input vector.
Diffusion model training: we convert the 1D input into a sequence of T tokens, by applying linear embeddings to each voxel patch. The DiT model is trained on these token sequences while conditioning on covariates like Age, Sex, Bv, Gf, Edu, BMI, etc. The training process involves forward diffusion and reverse denoising to model the distribution of voxel-level measures.
Diffusion model evaluation: we use the Metrics mean absolute error (MAE), root mean square error (RMSE) and calculate generated samples' X-Y Pearson correlation coefficient to evaluate the DiT model's performance
Normative Modeling and Z-score Calculation: Normative Modeling and Z-score Calculation: If the data is from other sources, we will finetune the DiT model in the new dataset. Subsequently, we employ the DDPM strategy to sample N voxel-level measures by conditioning on each covariates group. After that, we calculate voxel-level Z-scores using the mean and variance for the purpose of further analysis.
Data preparation: we collect the healthy population data with complete labels and scans. In this study, we use the FSL's FMRIB58_FA template to extract voxel data from registered scans and organize them into a one-dimensional (1D) input vector.
Diffusion model training: we convert the 1D input into a sequence of T tokens, by applying linear embeddings to each voxel patch. The DiT model is trained on these token sequences while conditioning on covariates like Age, Sex, Bv, Gf, Edu, BMI, etc. The training process involves forward diffusion and reverse denoising to model the distribution of voxel-level measures.
Diffusion model evaluation: we use the Metrics mean absolute error (MAE), root mean square error (RMSE) and calculate generated samples' X-Y Pearson correlation coefficient to evaluate the DiT model's performance
Normative Modeling and Z-score Calculation: Normative Modeling and Z-score Calculation: If the data is from other sources, we will finetune the DiT model in the new dataset. Subsequently, we employ the DDPM strategy to sample N voxel-level measures by conditioning on each covariates group. After that, we calculate voxel-level Z-scores using the mean and variance for the purpose of further analysis.
·Figure 1. The framework of generative voxel-level normative modeling method.
Results:
In our experiments, we used DWI FA scans as the neuroimaging measures. Age, Sex, Bv, Gf, Edu, BMI these six covariates were used as DiT's conditions. We selected widely adopted GAMLSS (Dinga, 2021), MFPR (Ge, 2024), diffusion model U-ViT (Bao, 2023) as comparative methods. Following the benchmark (Ge, 2024), we evaluated models' performance using mean absolute rrror (MAE) and root mean squared error (RMSE) on 1000 test cases. DiT achieved MAE and RMSE scores of 0.071 and 0.095, which were lowest among all comparative models. Additionally, we calculated measure-covariate Pearson correlaton coefficient vectors on Age, Sex, Bv for generated samples, training samples, and then calculate the cosine similarity between these vectors. The final similarity scores are 0.967, 0.940 and 0.794, which means that the generated samples effectively preserve the conditional information.
·Figure 2. Illustrative examples of comparative methods performance on 1000 test cases.
Conclusions:
The proposed voxel-level normative modeling method leverages the powerful generative diffusion models DiT to achieve precise voxel-level assessments. This method demonstrates generalizability to previously unseen samples during training, outperforming existing baseline methods in prediction accuracy and covariate correlation. In the future, we will extend our model to more imaging modalities and broader range of psychiatric disorders diagnosis.
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 2
Methods Development 1
Keywords:
Computing
Data analysis
Machine Learning
Modeling
MRI
Psychiatric Disorders
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
Dhariwal, P. (2021). Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34, 8780-8794.
Dinga, R. (2021). Normative modeling of neuroimaging data using generalized additive models of location scale and shape. BioRxiv, 2021-06.
Ge, R. (2024). Normative modelling of brain morphometry across the lifespan with CentileBrain: Algorithm benchmarking and model optimisation. The Lancet Digital Health, 6(3), e211-e221.
Marquand, A. F. (2016). Understanding heterogeneity in clinical cohorts using normative models: Beyond case-control studies. Biological Psychiatry, 80(7), 552-561.
Peebles, W. (2023). Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 4195-4205).
Rutherford, S. (2022). The normative modeling framework for computational psychiatry. Nature Protocols, 17(7), 1711-1734.