Diffusion-Driven Generation of Minimally Preprocessed Brain MRI

2025-10-29

AI Summary

Existing 3D denoising diffusion probabilistic models (DDPMs) for brain MRI typically require preprocessing—such as skull-stripping, spatial registration, and latent-space compression—limiting anatomical fidelity and generalizability to real-world imaging variability. Method: We introduce the first 3D DDPM trained directly on raw, unprocessed, unregistered T1-weighted brain MRI scans—including inherent anatomical diversity and magnetic field inhomogeneities—without latent-space compression. We benchmark three DDPM variants—score-based, velocity-based, and flow-based—on a large-scale dataset of 82,000 scans from BraTS and ADNI. Contribution/Results: Quantitative evaluation (FID, segmentation volume distributions) and qualitative analysis demonstrate that all models generate anatomically coherent 3D brain volumes. Velocity- and flow-based models achieve significantly lower FID (−12.4%) and better match ground-truth thalamic and lentiform nucleus volume distributions (KL divergence reduced by 0.18). All model weights and inference code are publicly released.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: (Large) Language Models

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

The purpose of this study is to present and compare three denoising diffusion probabilistic models (DDPMs) that generate 3D $T_1$-weighted MRI human brain images. Three DDPMs were trained using 80,675 image volumes from 42,406 subjects spanning 38 publicly available brain MRI datasets. These images had approximately 1 mm isotropic resolution and were manually inspected by three human experts to exclude those with poor quality, field-of-view issues, and excessive pathology. The images were minimally preprocessed to preserve the visual variability of the data. Furthermore, to enable the DDPMs to produce images with natural orientation variations and inhomogeneity, the images were neither registered to a common coordinate system nor bias field corrected. Evaluations included segmentation, Frechet Inception Distance (FID), and qualitative inspection. Regarding results, all three DDPMs generated coherent MR brain volumes. The velocity and flow prediction models achieved lower FIDs than the sample prediction model. However, all three models had higher FIDs compared to real images across multiple cohorts. In a permutation experiment, the generated brain regional volume distributions differed statistically from real data. However, the velocity and flow prediction models had fewer statistically different volume distributions in the thalamus and putamen. In conclusion this work presents and releases the first 3D non-latent diffusion model for brain data without skullstripping or registration. Despite the negative results in statistical testing, the presented DDPMs are capable of generating high-resolution 3D $T_1$-weighted brain images. All model weights and corresponding inference code are publicly available at https://github.com/piksl-research/medforj .

Problem

• Generate 3D T1-weighted brain MRI images using diffusion models
• Compare three diffusion models trained on minimally preprocessed data
• Evaluate generated image quality without skull-stripping or registration

Innovation

• Uses diffusion models for 3D brain MRI generation
• Employs minimal preprocessing without registration or bias correction
• First non-latent 3D diffusion model without skullstripping