MedForj: An open, large-scale foundational generative prior for high-resolution 3D brain MRI

Samuel W. Remedios; Aaron Carass; Jerry L. Prince; Blake E. Dewey

eess.IVOct 29, 2025

MedForj: An open, large-scale foundational generative prior for high-resolution 3D brain MRI

Authors: Samuel W. Remedios, Aaron Carass, Jerry L. Prince, Blake E. Dewey

Organizations: Image Analysis and Communications Laboratory, Johns Hopkins University · Department of Neurology, Johns Hopkins Hospital

Abstract

This work introduces MedForj, a suite of 3D foundational generative priors based on diffusion models. The MedForj models were trained on $72{,}659$ 1~mm isotropic 3D $T_1$ -weighted MRI human brain image volumes from $38{,}174$ subjects, drawn from a curated corpus of $80{,}675$ volumes from $42{,}506$ subjects spanning $38$ publicly available datasets. These training images were manually inspected to exclude those with poor quality and excessive pathology, and otherwise were minimally processed. The models include six different diffusion training strategies: rectified flow, latent diffusion rectified flow, flow matching, velocity prediction, clean prediction, and noise prediction. Image samples produced by each of these models were compared to each other and against real, ground truth data under downstream segmentation distributions, FID, five inverse problems, and blind human inspection in an observer study. Flow matching was the strongest strategy overall, achieving the best inverse problem solving results at $28.80$ ~dB PSNR and $0.874$ SSIM averaged over the five forward problems, the highest rate of reconstructions judged real by blind human raters at $72.6\%$ , and the closest per-structure match to real segmented anatomy in a permutation test. It was not best everywhere: rectified flow produced the most convincing unconditional samples in the observer study and the best FID, and the latent rectified-flow model achieved the smallest joint distributional distance to real anatomy. No other strategy, however, performed consistently well across all four evaluations. We therefore recommend flow matching as the default MedForj prior, while releasing every strategy so that the choice can be revisited per application. All model weights and corresponding code are publicly available at https://github.com/piksl-research/medforj.

Figures & tables

Dataset Name	Subjects (S)	Volumes (V)	Train S	Train V	Test S	Test V
A4	1742	6903	1594	6320	148	583
ABIDE	781	781	715	715	66	66
ABIDEII	959	1221	878	1112	81	109
ADNI	2587	17965	2368	16505	219	1460
AIBL Ellis et al. (2009)	673	1203	—	—	673	1203
AOMIC-ID1000 Snoek et al. (2021)	928	2768	850	2539	78	229

Table 1 : Distribution of subjects (S) and image volumes (V) by dataset. Note that this is not the total number of data available per dataset, but rather the amount of data after the exclusion described in Sec. 4.1 and Sec. 4.2 .

Figure 1 : A sagittal slice of a representative image used in the inverse problem experiments, with each row corresponding to a different inverse problem. The first column is identical across all rows, showing the ground truth. The second column shows the degraded input for each method. Following columns show the result of each strategy for solving the inverse problem. Yellow boxes indicate a zoomed region, shown in Fig. 2 .

Figure 2 : Zoomed regions corresponding to the yellow boxes in Fig. 1 . Rows and columns follow the same layout: one inverse problem per row, ground truth and degraded input in the first two columns, and one training strategy per remaining column.

PSNR (dB) $\uparrow$
Forward Problem	Input	Noise	Clean	Vel.	Flow	R. Flow	LDM
Super-resolution	21.95	13.47	24.53	24.02	24.77	25.17 ∗	23.09
Inpainting	26.83	20.73	31.64	26.07	32.76 ∗	28.06	25.11
Rician denoising	18.90	25.53	25.68	21.38	25.24	26.97 ∗	17.76
$k$ -space accel.	19.25	16.36	27.91	24.87	28.10 ∗	27.03	22.88
Motion	21.76	21.31	31.97	25.40	33.14 ∗	29.84	25.82

Table 2: Quantitative results (PSNR, SSIM, LPIPS) for the inverse problem solving task over

100

subjects. The final row for each metric shows the average score across all tasks for that metric. “Input” is the input image for each inverse problem, except for super-resolution whose images are linearly interpolated. Other abbreviations: “Vel.” is velocity prediction, “R. Flow” is rectified flow, and “LDM” is the latent diffusion model rectified flow. Bolded numbers indicate the best scoring strategy for that task, and asterisks ∗ indicate statistical significance (

\alpha=0.05

) to a two-sided paired Wilcoxon signed-rank test Wilcoxon (1945) after Holm-Bonferroni Holm (1979) correction.

Figure 3 : Biplanar views of three uncurated samples from each training strategy. These three samples share the exact random seed across all training strategies. All volumes were generated with 256 DDIM steps.

Figure 4 : Biplanar views of synthetic images and their two nearest neighbors. Nearest neighbor is computed as the minimum mean squared error in voxel space.

	Denoising steps (NFE)
Strategy	16	32	64	128	256	512	1000
Noise ( $\epsilon$ ) ‡	131.91	135.34	141.09	144.74	142.12	140.54	143.70
Clean ( $x_{0}$ )	58.07 †	55.17	52.24	54.63	56.76	58.99 †	58.40 †
Velocity	47.23 †	42.41 †	37.43	37.46	35.99	37.19	37.10
Flow	36.74 †	29.95 †	23.44 †	22.54 †	21.50	20.86	20.43
Rectified flow	59.51 †	34.55 †	23.36 †	16.41 †	15.52	14.27	14.38

Table 3: Ablation study to select the number of denoising steps, using FID between

100

samples from each strategy and the entire internal training corpus. Bolded cells give each strategy’s best (lowest) FID. Daggered † cells exceed that strategy’s own best FID by more than

10\%

and so fail the selection rule at that step count. The operating point is the fewest steps at which every strategy is within

10\%

of its own best FID; this is

256

, identically against all three real references. Noise prediction ‡ is reported for completeness but excluded from the rule, as it failed to converge and its FID never approaches that of the other strategies. A stricter

5\%

bound was pre-registered first and returned no step count common to all strategies, so the

10\%

bound reported here is the fallback specified by the same pre-registration rather than a threshold chosen after seeing the data.

Strategy	WASABI ( $\times 10^{4}$ ) $\downarrow$	SD	$\times$ floor
Real vs. real	0.54	0.12	—
Noise ( $\epsilon$ )	32.77	5.14	60.7
Clean ( $x_{0}$ )	6.09	0.82	11.3
Velocity	5.70	0.71	10.6
Flow	4.20	0.78	7.8
Rectified flow	9.99	1.25	18.5

Table 5 : The WASABI score for each training strategy. SD is the standard deviation of the WASABI score.

\times

floor is the row’s distance divided by the real vs. real distance in the first row.

Figure 5 : Raincloud plots showing the distributions of volumes from six regions of interest as computed by SynthSeg. Horizontal lines extend from the real distribution quartiles to aid comparison across strategies.

Mean drift (%) — positive is larger than real
Structure	Noise	Clean	Velocity	Flow	Rectified flow	LDM
Cerebral cortex	$-2.2$	$+11.6$	$+12.4$	$+2.0$	$-0.5$	$-1.9$
Brain stem	$-33.2$	$+15.9$	$+20.3$	$+5.7$	$-0.4$	$+1.1$
Ventricles	$-13.6$	$+28.8$	$+30.8$	$+29.3$	$-21.6$	$-14.9$
Thalamus	$-44.3$	$+6.9$	$+11.1$	$+0.2$	$+0.4$	$+0.1$
Putamen	$-48.9$	$-0.7$	$+8.9$	$+0.1$	$+0.1$	$-4.1$

Table 6 : Per-structure drift of the synthetic SynthSeg volume distributions from real, for all six strategies. Mean drift is

100\times(\bar{x}_{\mathrm{model}}-\bar{x}_{\mathrm{real}})/\bar{x}_{\mathrm{real}}

, so positive means the synthetic structure is larger than real. Cohen’s

d

restates that difference in pooled-SD units so structures of very different absolute size can be compared. SD ratio is the model’s standard deviation divided by the real standard deviation for that structure:

1.0

reproduces the real spread, below

1.0

means the model generates less anatomical volume variation than real, above

1.0

means more.

	Judged real (%, [95% CI])
Strategy	Generative samples	Reconstructions
	Study A	Study B
Real images	79.5 [76.4, 82.3]	—
Noise ( $\epsilon$ )	4.2 [2.0, 8.3]	24.4 [19.0, 30.7]
Clean ( $x_{0}$ )	31.0 [24.2, 38.6]	70.5 [63.7, 76.6]
Velocity	40.6 [33.7, 47.9]	46.2 [38.9, 53.7]

Table 8: Blind observer study. Percentage of lightboxes judged real by human raters who were not told which method produced any image, with Wilson 95% confidence intervals. 2949 reads from 7 independent raters, none of them authors. Raters saw one lightbox at a time and judged it real or generated; an unsure response is counted as not real throughout, so a lightbox counts as passing for real only when a rater affirmatively said so. Generative samples (Study A) are unconditional samples, where a higher rate means the sample passed for real. Reconstructions (Study B) are solutions to the inverse problems; these have no correct real/generated label — a reconstruction of a real subject is neither — so they are reported only as the rate judged real and never as an accuracy. Real images is the study’s ceiling. Bold marks the highest rate within a column.

Panel of Figure 6 — Figure 6 : Parasagittal slices from representative pass and fail images. (a) Images that passed manual quality assurance. (b) Images that failed manual quality assurance. Failure reasons include Pathology ( $\bullet$ ), Limited field-of-view (FOV) ( $\blacksquare$ ), Aggressive defacing ( $\blacktriangle$ ), Motion artifacts ( $\blacksquare$ ), and Noise ( $\blacktriangle$ ).

Explore similar work

VoxStruct3D: Structure-Leading Flow Matching for Voxel-Space 3D MRI Synthesis
Aug 5, 2026Fang Li, Yang Gao, Shihao Zou +5Voxel Medical Image Generation
State Key Laboratory of Virtual Reality Technology and Systems, Beihang University · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Faculty of Computer Science and Control Engineering, Shenzhen University of Advanced Technology
MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models
Aug 10, 2026Lisa K. Fischer, Mykhailo Riabets, Daniel Rueckert +3Magnetic Resonance Imaging Dataset Medical Image Generation
Technical University of Munich (TUM), Germany · Chair for AI in Healthcare and Medicine, TUM and TUM University Hospital, Munich, Germany · Imperial College London +4
AnaDiffusion: Anatomically CompositionalLatent Diffusion for Controllable 3D Brain MRI Generation
Aug 24, 2026Huiwen Han, Lulin Liu, Bangya Liu +8Magnetic Resonance Imaging Identity-Conditioned Latent Diffusion Models
Stanford University · Texas A&M University · University of Minnesota +4

Aug 5, 2026, cs.CVJ/K move · Enter open · S save

VoxStruct3D: Structure-Leading Flow Matching for Voxel-Space 3D MRI Synthesis

Fang Li, Yang Gao, Shihao Zou, Weixin Si+4

State Key Laboratory of Virtual Reality Technology and Systems, Beihang University · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Faculty of Computer Science and Control Engineering, Shenzhen University of Advanced Technology

Voxel Medical Image Generation 3D Generative Models 3D Generation Autoencoder Architectures

Open paper AskPDF arXiv

High-fidelity 3D MRI synthesis requires both globally coherent anatomy and fine-grained voxel-level detail. Although latent diffusion makes volumetric generation tractable, its image autoencoder introduces a reconstruction bottleneck that can limit the fine detail recoverable in the final volume. We present VoxStruct3D, a voxel-space flow-matching framework that directly models full-resolution MRI volumes using a clean-data prediction objective. Its Volumetric Voxel Generator (VVG) combines factorized 3D patch embedding with overlapping upsampling, time-modulated residual refinement, and skip fusion, enabling neighboring tokens to jointly reconstruct shared voxel regions and suppress patch-boundary artifacts. To complement direct voxel-space modeling with an explicit anatomical prior, we further introduce a Structure-First, Image-Follows (SFIF) strategy. A frozen pretrained 3D medical encoder and a StructVAE extract compact structure tokens that preserve dominant anatomy, while a structure-leading schedule keeps their trajectory ahead of the image trajectory. Patch-Aligned RoPE spatially aligns the unequal token grids, and asymmetric attention enforces one-way guidance from structure to image. Experiments on pathological and healthy T1-weighted brain MRI datasets show that VoxStruct3D achieves the strongest overall performance across feature-distribution alignment, sample diversity, and perceptual quality, producing anatomically coherent and visually realistic volumes.

Abstract

Explore similar work

VoxStruct3D: Structure-Leading Flow Matching for Voxel-Space 3D MRI Synthesis

MRIComp4Flow: Compression of 3D Brain MRI for Training Multi-Modal Generative Models

AnaDiffusion: Anatomically CompositionalLatent Diffusion for Controllable 3D Brain MRI Generation