Diffusion MRI

Diffusion MRI (DTI)

Join the Competition Contribute to Benchmark

Standard reconstruction benchmark — forward model perfectly known, no calibration needed. Score = 0.5 × clip((PSNR−15)/30, 0, 1) + 0.5 × SSIM

#	Method	Score	PSNR (dB)	SSIM	Source
🥇	DiffusionDTI DiffusionDTI Gao et al., NeurIPS 2024 39.1 dB SSIM 0.952 Checkpoint unavailable	0.878	39.1	0.952	✓ Certified	Gao et al., NeurIPS 2024
🥈	PhysDiffMRI PhysDiffMRI Chen et al., MRM 2024 37.5 dB SSIM 0.941 Checkpoint unavailable	0.845	37.5	0.941	✓ Certified	Chen et al., MRM 2024
🥉	SwinDTI SwinDTI Wang et al., MICCAI 2023 36.2 dB SSIM 0.931 Checkpoint unavailable	0.819	36.2	0.931	✓ Certified	Wang et al., MICCAI 2023
4	DTIFormer DTIFormer Liu et al., MICCAI 2022 34.8 dB SSIM 0.912 Checkpoint unavailable	0.786	34.8	0.912	✓ Certified	Liu et al., MICCAI 2022
5	DWIML-Net DWIML-Net Qin et al., IEEE TMI 2019 32.1 dB SSIM 0.871 Checkpoint unavailable	0.721	32.1	0.871	✓ Certified	Qin et al., IEEE TMI 2019
6	DnCNN-DTI DnCNN-DTI Golkov et al., IEEE TMI 2016 29.3 dB SSIM 0.831 Checkpoint unavailable	0.654	29.3	0.831	✓ Certified	Golkov et al., IEEE TMI 2016
7	CHARMED CHARMED Assaf & Basser, NeuroImage 2005 26.8 dB SSIM 0.782 Try in SpecLab →	0.588	26.8	0.782	✓ Certified	Assaf & Basser, NeuroImage 2005
8	SHORE SHORE Merlet & Deriche, MRM 2013 24.6 dB SSIM 0.745 Try in SpecLab →	0.532	24.6	0.745	✓ Certified	Merlet & Deriche, MRM 2013
9	DTI-FIT DTI-FIT Behrens et al., MRM 2003 22.4 dB SSIM 0.710 Try in SpecLab →	0.478	22.4	0.710	✓ Certified	Behrens et al., MRM 2003

Dataset: PWM Benchmark (9 algorithms)

Blind Reconstruction Challenge — forward model has unknown mismatch, must calibrate from data. Score = 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)

#	Method	Overall Score	Public PSNR / SSIM	Dev PSNR / SSIM	Hidden PSNR / SSIM	Trust	Source
🥇	DiffusionDTI + gradient DiffusionDTI + gradient Gao et al., NeurIPS 2024 Score 0.777 Correct & Reconstruct →	0.777	0.858 37.87 dB / 0.982	0.750 31.4 dB / 0.936	0.724 29.08 dB / 0.902	✓ Certified	Gao et al., NeurIPS 2024
🥈	PhysDiffMRI + gradient PhysDiffMRI + gradient Chen et al., MRM 2024 Score 0.752 Correct & Reconstruct →	0.752	0.818 35.69 dB / 0.972	0.743 30.15 dB / 0.919	0.696 27.77 dB / 0.876	✓ Certified	Chen et al., MRM 2024
🥉	SwinDTI + gradient SwinDTI + gradient Wang et al., MICCAI 2023 Score 0.725 Correct & Reconstruct →	0.725	0.802 34.43 dB / 0.964	0.724 28.85 dB / 0.898	0.650 26.34 dB / 0.841	✓ Certified	Wang et al., MICCAI 2023
4	DTIFormer + gradient DTIFormer + gradient Liu et al., MICCAI 2022 Score 0.723 Correct & Reconstruct →	0.723	0.783 32.87 dB / 0.951	0.713 28.48 dB / 0.891	0.672 27.23 dB / 0.864	✓ Certified	Liu et al., MICCAI 2022
5	DWIML-Net + gradient DWIML-Net + gradient Qin et al., IEEE TMI 2019 Score 0.662 Correct & Reconstruct →	0.662	0.764 30.61 dB / 0.926	0.635 25.0 dB / 0.802	0.588 23.08 dB / 0.734	✓ Certified	Qin et al., IEEE TMI 2019
6	DnCNN-DTI + gradient DnCNN-DTI + gradient Golkov et al., IEEE TMI 2016 Score 0.589 Correct & Reconstruct →	0.589	0.693 27.18 dB / 0.862	0.559 21.67 dB / 0.676	0.516 20.08 dB / 0.603	✓ Certified	Golkov et al., IEEE TMI 2016
7	SHORE + gradient SHORE + gradient Merlet & Deriche, MRM 2013 Score 0.576 Correct & Reconstruct →	0.576	0.614 23.16 dB / 0.737	0.577 22.65 dB / 0.717	0.536 21.5 dB / 0.668	✓ Certified	Merlet & Deriche, MRM 2013
8	CHARMED + gradient CHARMED + gradient Assaf & Basser, NeuroImage 2005 Score 0.500 Correct & Reconstruct →	0.500	0.668 25.48 dB / 0.817	0.435 17.84 dB / 0.492	0.397 16.48 dB / 0.425	✓ Certified	Assaf & Basser, NeuroImage 2005
9	DTI-FIT + gradient DTI-FIT + gradient Behrens et al., MRM 2003 Score 0.471 Correct & Reconstruct →	0.471	0.512 19.76 dB / 0.587	0.467 18.45 dB / 0.522	0.434 17.99 dB / 0.500	✓ Certified	Behrens et al., MRM 2003

Complete score requires all 3 tiers (Public + Dev + Hidden).

Join the competition →

Scoring: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖) PSNR 40% · SSIM 40% · Consistency 20%

Public 3 scenes

Full-access development tier with all data visible.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.

How to use: Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.

What to submit: Reconstructed signals (x_hat) and corrected spec as HDF5.

Public Leaderboard

#	Method	Score	PSNR	SSIM
1	DiffusionDTI + gradient	0.858	37.87	0.982
2	PhysDiffMRI + gradient	0.818	35.69	0.972
3	SwinDTI + gradient	0.802	34.43	0.964
4	DTIFormer + gradient	0.783	32.87	0.951
5	DWIML-Net + gradient	0.764	30.61	0.926
6	DnCNN-DTI + gradient	0.693	27.18	0.862
7	CHARMED + gradient	0.668	25.48	0.817
8	SHORE + gradient	0.614	23.16	0.737
9	DTI-FIT + gradient	0.512	19.76	0.587

Spec Ranges (4 parameters)

Parameter	Min	Max	Unit
b_value_error	-3.0	6.0	%
eddy_current	-0.5	1.0	voxels
gradient_direction	-1.0	2.0	deg
susceptibility	-1.0	2.0	voxels

View Details →

Dev 3 scenes

Blind evaluation tier — no ground truth available.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), and spec ranges only.

How to use: Apply your pipeline from the Public tier. Use consistency as self-check.

What to submit: Reconstructed signals and corrected spec. Scored server-side.

Dev Leaderboard

#	Method	Score	PSNR	SSIM
1	DiffusionDTI + gradient	0.750	31.4	0.936
2	PhysDiffMRI + gradient	0.743	30.15	0.919
3	SwinDTI + gradient	0.724	28.85	0.898
4	DTIFormer + gradient	0.713	28.48	0.891
5	DWIML-Net + gradient	0.635	25.0	0.802
6	SHORE + gradient	0.577	22.65	0.717
7	DnCNN-DTI + gradient	0.559	21.67	0.676
8	DTI-FIT + gradient	0.467	18.45	0.522
9	CHARMED + gradient	0.435	17.84	0.492

Spec Ranges (4 parameters)

Parameter	Min	Max	Unit
b_value_error	-3.6	5.4	%
eddy_current	-0.6	0.9	voxels
gradient_direction	-1.2	1.8	deg
susceptibility	-1.2	1.8	voxels

Submit Reconstruction View Details →

Hidden 3 scenes

Fully blind server-side evaluation — no data download.

What you get & how to use

What you get: No data downloadable. Algorithm runs server-side on hidden measurements.

How to use: Package algorithm as Docker container / Python script. Submit via link.

What to submit: Containerized algorithm accepting y + H, outputting x_hat + corrected spec.

Hidden Leaderboard

#	Method	Score	PSNR	SSIM
1	DiffusionDTI + gradient	0.724	29.08	0.902
2	PhysDiffMRI + gradient	0.696	27.77	0.876
3	DTIFormer + gradient	0.672	27.23	0.864
4	SwinDTI + gradient	0.650	26.34	0.841
5	DWIML-Net + gradient	0.588	23.08	0.734
6	SHORE + gradient	0.536	21.5	0.668
7	DnCNN-DTI + gradient	0.516	20.08	0.603
8	DTI-FIT + gradient	0.434	17.99	0.5
9	CHARMED + gradient	0.397	16.48	0.425

Spec Ranges (4 parameters)

Parameter	Min	Max	Unit
b_value_error	-2.1	6.9	%
eddy_current	-0.35	1.15	voxels
gradient_direction	-0.7	2.3	deg
susceptibility	-0.7	2.3	voxels

Submit Algorithm View Details →

Blind Reconstruction Challenge

Challenge

Given measurements with unknown mismatch and spec ranges (not exact params), reconstruct the original signal. A method must be evaluated on all three tiers for a complete score. Scored on a composite metric: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖).

Input

Measurements y, ideal forward model H, spec ranges

Output

Reconstructed signal x̂

About the Imaging Modality

Diffusion MRI measures the random Brownian motion of water molecules in tissue by applying magnetic field gradient pulses that encode microscopic displacement. The signal attenuation follows S = S_0 * exp(-b * D_eff) where b is the diffusion weighting factor and D_eff is the effective diffusion coefficient along the gradient direction. Acquiring measurements in multiple gradient directions enables estimation of the diffusion tensor (DTI) and derived scalar maps (FA, MD, AD, RD). Advanced models (NODDI, CSD) resolve intra-voxel fiber crossings. Primary degradations include EPI distortion, eddy currents, and motion sensitivity.

Principle

Diffusion MRI sensitizes the MR signal to the Brownian motion of water molecules by applying strong magnetic field gradient pulses (Stejskal-Tanner scheme). In fibrous tissue (e.g., white matter), water diffuses preferentially along fibers, creating directional diffusion anisotropy. Diffusion Tensor Imaging (DTI) models this as a 3×3 tensor; higher-order models (HARDI, CSD) resolve crossing fibers.

How to Build the System

Acquire on a 3T scanner with high-performance gradients (80 mT/m, 200 T/m/s). Use spin-echo EPI with multiple b-values (e.g., b=0, 1000, 2000 s/mm²) and 30-300 diffusion directions uniformly distributed on the sphere. Include reverse-phase-encode b=0 images for EPI distortion correction. Multi-band (SMS) acceleration reduces scan time. Typical parameters: 2 mm isotropic, TE 60-90 ms, TR 3-5 s.

Common Reconstruction Algorithms

DTI tensor fitting (least-squares or weighted least-squares)
CSD (Constrained Spherical Deconvolution) for fiber orientation distribution
NODDI (Neurite Orientation Dispersion and Density Imaging)
Probabilistic tractography (FSL probtrackx, MRtrix3 iFOD2)
Deep-learning tract segmentation (TractSeg, DeepBundle)

Common Mistakes

Eddy current and EPI geometric distortions not corrected, causing tract errors
Insufficient number of diffusion directions for the chosen model complexity
Using DTI in regions with crossing fibers, producing incorrect FA and tract directions
Susceptibility-induced signal dropout near air-tissue interfaces (sinuses, temporal lobes)
Head motion between diffusion volumes causing inter-volume misalignment

How to Avoid Mistakes

Apply FSL eddy or equivalent for eddy current, motion, and susceptibility correction
Use ≥30 directions for DTI, ≥60 for CSD, and ≥90 for multi-shell models
Use multi-fiber models (CSD, NODDI) in regions known to have crossing fibers
Use reduced FOV or multi-shot EPI near susceptibility-prone regions
Include interspersed b=0 volumes for robust motion and drift correction

Forward-Model Mismatch Cases

The widefield fallback produces a blurred spatial image, but diffusion MRI applies magnetic field gradients to encode Brownian water motion — the Stejskal-Tanner signal attenuation S = S_0*exp(-b*D) is not modeled
Diffusion MRI acquires multiple volumes at different b-values and gradient directions to measure the diffusion tensor at each voxel — the widefield single-image model cannot encode directional water diffusivity or fiber orientation

How to Correct the Mismatch

Use the diffusion MRI operator that applies Stejskal-Tanner encoding: y_i = FFT(x * exp(-b_i * g_i^T * D * g_i)) for each gradient direction g_i and b-value b_i
Reconstruct diffusion tensors (DTI) or fiber orientation distributions (CSD, NODDI) from the multi-direction, multi-b-value measurements using the correct diffusion-weighted forward model

Experimental Setup — Signal Chain

Experimental setup diagram for Diffusion MRI (DTI)

Reconstruction Gallery — 4 Scenes × 3 Scenarios

Method: CPU_baseline | Mismatch: nominal (nominal=True, perturbed=False)

I Ideal II Mismatch (uncorrected) III Oracle correction

I — Ideal

Ground Truth

Measurement

Reconstruction

16.584200964385797 dB SSIM 0.3250905580130585

II — Mismatch

Ground Truth

Measurement

Reconstruction

12.506523882319042 dB SSIM 0.044716474341606735

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

15.931719256055723 dB SSIM 0.12142227244746187

I — Ideal

Ground Truth

Measurement

Reconstruction

18.884299556167566 dB SSIM 0.3252411352628261

II — Mismatch

Ground Truth

Measurement

Reconstruction

13.742128073130111 dB SSIM 0.048029248758291385

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

17.02391116326154 dB SSIM 0.11622117666622035

I — Ideal

Ground Truth

Measurement

Reconstruction

18.301830801494646 dB SSIM 0.32976160086547746

II — Mismatch

Ground Truth

Measurement

Reconstruction

13.36792873190369 dB SSIM 0.04964364838717723

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

16.550610792224443 dB SSIM 0.1259030967011571

I — Ideal

Ground Truth

Measurement

Reconstruction

16.67775192548055 dB SSIM 0.3247822274360334

II — Mismatch

Ground Truth

Measurement

Reconstruction

12.53195005733068 dB SSIM 0.044966709218701925

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

15.850507045662987 dB SSIM 0.12621817875162467

Mean PSNR Across All Scenes

17.61202081188214 dB

I — Ideal

13.037132686170882 dB

II — Mismatch

16.33918706430117 dB

III — Oracle

Degradation: -4.6 dB Recovery: +3.3 dB

Per-scene PSNR breakdown (4 scenes)

Scene	I (PSNR)	I (SSIM)	II (PSNR)	II (SSIM)	III (PSNR)	III (SSIM)
scene_00	16.584200964385797	0.3250905580130585	12.506523882319042	0.044716474341606735	15.931719256055723	0.12142227244746187
scene_01	18.884299556167566	0.3252411352628261	13.742128073130111	0.048029248758291385	17.02391116326154	0.11622117666622035
scene_02	18.301830801494646	0.32976160086547746	13.36792873190369	0.04964364838717723	16.550610792224443	0.1259030967011571
scene_03	16.67775192548055	0.3247822274360334	12.53195005733068	0.044966709218701925	15.850507045662987	0.12621817875162467
Mean	17.61202081188214	0.3262188803943489	13.037132686170882	0.046839020176444326	16.33918706430117	0.122441181141616

Experimental Setup

Instrument: Siemens MAGNETOM Prisma (HCP protocol)

Field Strength T: 3.0

B Values: [0, 1000]

Diffusion Directions: 64

Matrix Size: 128x128

Tr Ms: 4000

Te Ms: 80

Voxel Size Mm: 2x2x2

Sequence: single-shot spin-echo EPI DWI

Reconstruction: weighted least squares / RESTORE

Dataset: HCP, UK Biobank

Key References

Basser et al., 'MR diffusion tensor spectroscopy and imaging', Biophysical Journal 66, 259-267 (1994)
Sotiropoulos et al., 'Advances in diffusion MRI acquisition and processing in the HCP', NeuroImage 80, 125-143 (2013)

Canonical Datasets

Human Connectome Project (HCP) diffusion data
UK Biobank diffusion imaging

Spec DAG — Forward Model Pipeline

F(EPI) → D(g, η₁)

F Diffusion-Weighted EPI (EPI)

D RF Coil Receiver (g, η₁)

Mismatch Parameters

Symbol	Parameter	Description	Perturbed
Δb	b_value_error	b-value calibration error (%)	3.0
ε	eddy_current	Eddy-current distortion (voxels)	0.5
Δĝ	gradient_direction	Gradient direction error (deg)	1.0
Δχ	susceptibility	Susceptibility distortion (voxels)	1.0

Credits System

40%

Platform Profit Pool

Revenue allocated to benchmark rewards

30%

Winner Share

Top algorithm receives from pool

$100

Min Withdrawal

Minimum payout threshold

Spec Primitives Reference (11 primitives)

P Propagation

Free-space or medium propagation kernel (Fresnel, Rayleigh-Sommerfeld).

M Mask / Modulation

Spatial or spatio-temporal amplitude modulation (coded aperture, SLM pattern).

Π Projection

Geometric projection operator (Radon transform, fan-beam, cone-beam).

F Fourier Sampling

Sampling in the Fourier / k-space domain (MRI, ptychography).

C Convolution

Shift-invariant convolution with a point-spread function (PSF).

Σ Summation / Integration

Summation along a physical dimension (spectral, temporal, angular).

D Detector

Sensor readout with gain g and noise model η (Gaussian, Poisson, mixed).

S Structured Illumination

Patterned illumination (block, Hadamard, random) applied to the scene.

W Wavelength Dispersion

Spectral dispersion element (prism, grating) with shift α and aperture a.

R Rotation / Motion

Sample or gantry rotation (CT, electron tomography).

Λ Wavelength Selection

Spectral filter or monochromator selecting a wavelength band.