Quantitative analysis of synthetic 3D MRI generated by latent diffusion models

1981 

Abstract Submission 

Howook Lee¹, Won Hee Lee¹

¹Kyung Hee University, Yongin, Republic of Korea

Howook Lee  
Kyung Hee University
Yongin, Republic of Korea

Won Hee Lee  
Kyung Hee University
Yongin, Republic of Korea

The scarcity of medical data, particularly in the field of brain imaging, has posed a significant challenge for the development and advancement of artificial intelligence (AI) in neuroscience. The use of synthetic data has the potential to address the challenge of medical data scarcity and improve the training and validation of AI models for various medical applications. Diffusion models (Ho et al., 2020), specifically latent diffusion models (LDMs) (Rombach et al., 2022), have emerged as powerful tools for generating high-quality synthetic data that closely resembles real-world data. However, the quality of synthetic MRI data generated by LDMs is still poorly understood. Here, we aim to generate 3D synthetic MRI data using LDMs and comprehensively compare it to real-world MRI data in terms of visual appearance, morphological accuracy, and its suitability for predicting brain age.

We used T1-weighted MRI scans from 652 healthy participants aged 18 to 88 from the Cambridge Centre for Ageing and Neuroscience (Taylor et al., 2017). After strict preprocessing quality control of imaging data, our main analysis comprised 630 participants. LDMs were trained using the preprocessed data, with age and sex as conditions to generate high-quality synthetic MRI data. To ensure comparability, we acquired 630 synthetic MRI data from the trained LDMs, and matched the age and sex distributions between the original and synthetic MRI datasets. We then compared the synthetic MRI data to the original data in terms of visual quality, morphological features, and brain age prediction accuracy. For visual similarity, we computed the Fréchet inception distance (FID) (Heusel et al., 2017) and structural similarity index measure (SSIM) scores (Wang et al., 2004). For morphological feature comparison, we extracted morphological features, namely regional measures of cortical thickness (n = 68), surface area (n = 68), and subcortical volume (n = 16), from both original and synthetic MRI data using FreeSurfer (Fischl et al., 2012), and then compared their correlations with age. We further calculated Fisher's z-values to compare the correlation coefficients between age and morphological features for original and synthetic data. We also computed the relative error between ROI feature values in original and synthetic data. For brain age prediction, we trained a support vector machine (SVM) model using original and synthetic data separately, and evaluated their performance on the same real test set. The models were evaluated using a nested 10-fold cross-validation. We quantified the performance of the models using mean absolute error (MAE) and the coefficient of determination (R²) between predicted brain age and chronological age.

Figure 1 shows that the synthetic MRI data generated by LDMs closely resembles real MRI data. We obtained the FIDs of 13.18, 23.62, and 23.76 in the axial, coronal, and sagittal planes for the original data and those of 17.19, 41.34, and 34.04 for the synthetic data. The SSIM values were 0.81 for the original data and 0.82 for the synthetic data. Figure 2 shows the distribution of Fisher's z values and the relative errors for each ROI, offering insights into the statistical differences in correlations and magnitudes of ROI features between the original and synthetic data. The brain age prediction model yielded average MAE and R² values of 6.35 and 0.81 for the original data and 9.96 and 0.55 for the synthetic data, respectively.

Our results collectively provide a comprehensive assessment of the synthetic MRI data generated by LDMs, covering visual, quantitative, and predictive aspects. The findings suggest a close resemblance between the original and synthetic MRI datasets, but also highlight some differences, particularly in the FID values and metrics of the brain age prediction model. We provided multifaceted assessment for understanding the strengths and limitations of LDM-generated synthetic MRI data.

Aging ²

Other Methods ¹

Aging

Machine Learning

Morphometrics

MRI

Other - Generative models;Latent diffusion models;Synthetic data;Brain Imaging