Toward Clinical Translation of Deep Learning-Based Intracranial Super-Resolution 4D Flow MRI Using Multi-Cohort Transfer Learning

Oliver Welin Odeback; Susanne Schnell; Patrick Winter; Johan Lundberg; Fredrik Piehl; Tobias Granberg; Alexander Fyrdahl; David Marlevi

cs.CVOct 26, 2026

Toward Clinical Translation of Deep Learning-Based Intracranial Super-Resolution 4D Flow MRI Using Multi-Cohort Transfer Learning

Authors: Oliver Welin Odeback, Susanne Schnell, Patrick Winter, Johan Lundberg, Fredrik Piehl, Tobias Granberg, Alexander Fyrdahl, David Marlevi

Organizations: Department of Molecular Medicine and Surgery, Karolinska Institutet, Stockholm, Sweden. · Department of Medical Physics, University of Greifswald, Greifswald, Germany. · Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden. · Department of Neuroradiology, Karolinska University Hospital, Stockholm, Sweden. ·

Abstract

We evaluated synthetic pretraining and multi-cohort transfer learning for deep learning-based super-resolution of intracranial 4D Flow MRI across heterogeneous neurologic cohorts. This mixed prospective-retrospective study analyzed 18 participants (12 healthy participants, 3 with moyamoya disease, and 3 with multiple sclerosis) with paired 1.2-mm and 0.6-mm 4D Flow MRI scans. A convolutional neural network pretrained on synthetic cerebrovascular data was fine-tuned using cohort-specific or pooled multi-cohort in vivo data. High-resolution-derived inputs enabled voxel-wise evaluation; native standard-resolution scans were used for angiographic and hemodynamic analyses. Synthetic pretraining reduced mean artery-wise velocity-normalized root mean square error (vNRMSE) versus interpolation (0.147 vs 0.166). Pooled transfer learning yielded the lowest mean vNRMSE (0.119 vs 0.127 for cohort-specific models and 0.147 for synthetic pretraining), with narrower 95% limits-of-agreement widths than synthetic pretraining across all test cases (0.109-0.222 vs 0.154-0.278 m/s). Super-resolved phase-contrast angiography achieved higher precision-recall area under the curve than native standard-resolution imaging in all cohorts (0.848-0.896 vs 0.486-0.805). These results indicate that pooled transfer learning improves cross-cohort reconstruction agreement and that super-resolution can improve vessel depiction, supporting further multi-center validation for cerebrovascular imaging.

Figures & tables

Figure 1: Methodological overview. Paired standard- and high-resolution four-dimensional flow magnetic resonance imaging (4D Flow MRI) scans were acquired in healthy volunteers and patients ( $n=18$ ). A synthetic-pretrained convolutional neural network (CNN) was fine-tuned using in vivo data via transfer learning to super-resolve standard-resolution data, which were evaluated against the corresponding high-resolution reference scans.

	Overall	Intracranial	Extracranial	Neuro- vascular	Neuro- inflammatory
Participants, n	18	6	6	3	3
Age, y ∗	35.4 (23–70)	35.8 (25–70)	30.2 (23–38)	36.3 (31–41)	44.0 (34–59)
Women, n	8	2	2	3	1
Men, n	10	4	4	0	2
BMI ${}^{\dagger\,\ddagger}$	22.6 (19.8–32.3)	24.4 (19.8–28.7)	22.5 (20.6–32.3)	22.5 (21.3–24.1)	21.4 (21.2–24.8)
Heart rate, beats/min ‡	60.8 (49.3–71.5)	60.8 (50.0–68.3)	58.2 (49.3–61.8)	67.0 (60.0–71.5)	61.5 (60.0–65.7)

Table 1: Characteristics of study cohorts.

	Site 1		Site 2
Parameter	LR	HR	LR	HR
Resolution (mm)	1.2	0.6	1.2	0.6
TE (ms)	4.5	4.96	3.08	3.72
TR (ms)	100.8	103.6	79.8	92.4
Cardiac phases	5–8	5–8	11–12	9–11
Flip angle ( ∘ )	15	15	15	15

Table 2: MRI acquisition parameters.

Figure 2: Artery-wise voxel-wise reconstruction error relative to high-resolution (HR) 4D Flow MRI, quantified using velocity-normalized root mean square error (vNRMSE) and averaged over the full cardiac cycle. Results are shown for bicubic interpolation (Int.), the CFD-pretrained baseline (CFD), cohort-specific fine-tuned models (IC, EC, NV, NI), and the pooled multi-cohort model (POOL) across all 4D Flow MRI test sets. Yellow shaded regions indicate the cohort-specific fine-tuned model corresponding to the test domain in each panel. Green shaded regions indicate POOL. Point colors denote labeled arteries.

Figure 3: Voxel-wise Bland–Altman analysis between high-resolution (HR) reference velocities ( $V_{\mathrm{HR}}$ ) and super-resolved velocities ( $V_{\mathrm{SR}}$ ) at peak systole. Rows show 4DFlowNet Ferdian et al. (2023) , the CFD-pretrained baseline (CFD), the cohort-specific fine-tuned model (COHORT), and the pooled fine-tuned model (POOL); columns show the four test cases. For the intracranial, neurovascular, and neuroinflammatory cases, signed velocity components ( $V_{x},V_{y},V_{z}$ ) were pooled. For the extracranial case, only $V_{z}$ was analyzed, representing the predominant velocity component. Differences were defined as $V_{\mathrm{SR}}-V_{\mathrm{HR}}$ and plotted against the corresponding pairwise mean. Solid horizontal lines indicate mean bias and dashed lines the 95% limits of agreement (mean bias $\pm 1.96$ standard deviation), calculated using all analyzed voxels within the vessel mask. For visualization, an equal number of voxels per artery was sampled. Points are colored according to artery label.

Figure 4: Derived phase-contrast MR angiography (PCMRA) for the test subjects from each cohort. Rows show time-of-flight (TOF) angiography as anatomical reference, with PCMRA derived from acquired standard-resolution (LR) and high-resolution (HR) 4D Flow MRI, and PCMRA from the CFD-pretrained baseline (CFD), cohort-specific fine-tuned (COHORT), and pooled fine-tuned (POOL) super-resolution models. For each PCMRA image, the area under the precision–recall curve (PR-AUC) and maximum Dice coefficient (DICE) with respect to the corresponding TOF-derived vessel mask are reported. The highest PR-AUC and $Dice_{\max}$ values for each cohort are underlined.

Figure 5: Plane-based hemodynamic comparison across all test subjects. For each cohort, the time-of-flight angiogram indicates the analyzed vessel regions, with colored markers denoting the first and last analysis plane of selected plane groups. Velocity vector fields (Intracranial, Neurovascular) or streamlines (Extracranial, Neuroinflammatory) at peak systole are shown for standard-resolution (LR), high-resolution (HR), CFD-pretrained (CFD), and pooled fine-tuned (POOL) within the highlighted region. For visual depiction of SR reconstructions, the displayed fluid region was defined using the corresponding SR-derived PCMRA mask. Flow rate (Q, mL/s) and 95th-percentile velocity (v95, m/s) are shown over the cardiac cycle for each selected plane group, generated using TOF-derived masks across LR, HR, and super-resolved (SR) reconstructions. In the neurovascular case, four additional moyamoya-associated vessels (MMV-1–MMV-4) were analyzed within the disease-relevant region. Because these vessels were incompletely captured by TOF-derived arterial labels, super-resolved PCMRA-derived masks were used for both visualization and quantification of these vessels across LR, HR, and SR reconstructions.

Artery	LR	HR	CFD	POOL
Intracranial
RICA	5.25 (4.83, 5.76)	5.73 (5.39, 6.04)	5.45 (4.99, 5.81)	5.93 (5.38, 6.50)
LICA	3.73 (3.34, 4.29)	4.28 (3.92, 4.67)	4.02 (3.77, 4.41)	4.49 (4.24, 4.94)
BA	3.38 (3.12, 3.87)	4.24 (3.98, 4.38)	3.64 (3.52, 3.94)	4.10 (3.87, 4.46)
RMCA	2.78 (2.66, 2.99)	2.79 (2.22, 3.02)	3.00 (2.98, 3.09)	3.51 (3.34, 3.67)
LMCA	1.56 (1.31, 2.03)	0.98 (0.36, 1.41)	1.69 (1.63, 1.90)	2.61 (2.45, 2.95)

Table 3: Artery-level peak-systolic flow (

Q_{\mathrm{peak}}

, mL/s), summarized as median (Q1, Q3) across all analysis planes along each vessel centerline. Values are shown for standard-resolution (LR), high-resolution (HR), CFD-pretrained (CFD), and pooled fine-tuned (POOL) reconstructions. Quantification was performed using TOF-derived masks for all vessels except the moyamoya-associated vessels (MMV-1–MMV-4), which were analyzed using super-resolved PCMRA-derived masks because they were incompletely captured by the TOF-derived arterial labels. The largest median flow per artery is shown in bold; boldface is descriptive and does not indicate statistical significance. Note that in the neurovascular subject, the LMCA is not reported; the affected proximal M1 segment was instead represented by the moyamoya-associated vessels.

Figure 6: Voxel-wise peak-systolic velocity magnitude distributions, shown as violin plots for each labeled artery and cohort. The violin plots compare standard-resolution (LR), high-resolution (HR), CFD-pretrained (CFD), and pooled fine-tuned (POOL) reconstructions. Box plots within each violin indicate the median and interquartile range. For the primary vessel analyses, voxel-wise distributions were computed within TOF-derived arterial masks, shown as vessel renderings on the left. In the neurovascular case, four additional moyamoya-associated vessels (MMV-1–MMV-4) were analyzed within the disease-relevant region using super-resolved PCMRA-derived masks, shown in the zoomed vessel rendering, because these vessels were incompletely captured by the TOF-derived arterial labels.

	LR			HR			CFD			POOL
Artery	$v_{\mathrm{median}}$ (Q1, Q3)	$v_{\mathrm{p95}}$	div	$v_{\mathrm{median}}$ (Q1, Q3)	$v_{\mathrm{p95}}$	div	$v_{\mathrm{median}}$ (Q1, Q3)	$v_{\mathrm{p95}}$	div	$v_{\mathrm{median}}$ (Q1, Q3)	$v_{\mathrm{p95}}$	div
Intracranial
RICA	33.0 (20.9, 43.3)	60.8	6.9	38.6 (26.4, 50.6)	72.1	16.2	32.9 (16.7, 43.6)	61.3	3.3	33.0 (22.5, 43.0)	58.5	4.7
LICA	32.2 (20.0, 44.7)	62.4	6.8	39.3 (28.5, 50.9)	70.7	17.6	32.2 (16.4, 45.2)	62.4	3.5	33.1 (23.4, 44.1)	60.5	4.9
BA	31.7 (17.2, 40.8)	48.7	5.2	37.5 (25.8, 48.1)	61.5	18.5	31.9 (12.4, 41.9)	50.0	3.0	31.2 (21.1, 39.3)	47.4	3.7
RMCA	40.0 (16.5, 56.5)	71.8	6.4	58.2 (40.3, 72.5)	103.6	46.8	23.5 (2.3, 52.9)	69.0	4.9	32.4 (15.7, 51.5)	67.6	5.6

Table 4: Artery-level velocity distribution metrics at peak systole. Median velocity (

v_{\mathrm{median}}

), reported with the first and third quartiles (Q1, Q3), and the 95th percentile (

v_{\mathrm{p95}}

) characterize the voxel-wise velocity distribution within each artery, while mean absolute divergence (div) is reported in units of

10

s -1 . Values are shown for standard-resolution (LR), high-resolution (HR), CFD-pretrained (CFD), and pooled fine-tuned (POOL) reconstructions. Velocity values are given in cm/s. Quantification was performed using TOF-derived masks for all vessels except the moyamoya-associated vessels (MMV-1–MMV-4), which were analyzed using super-resolved PCMRA-derived masks because they were incompletely captured by the TOF-derived arterial labels. For readability, the largest value for each velocity-distribution metric and the lowest divergence value per artery are shown in bold; boldface is descriptive and does not indicate statistical significance. Note that in the neurovascular subject, the LMCA is not reported; the affected proximal M1 segment was instead represented by the moyamoya-associated vessels.

Figure S1 : Direction-consistency mask refinement example. Velocity vectors are shown for two representative, high-resolution (HR) test sets (Intracranial and Neurovascular) after refinement (PRUNED) and before (ORIGINAL), illustrating suppression of isolated locally incoherent voxels.

Cohort	Total, n	Training	Validation	Test
Intracranial healthy	6	4 participants	1 participant	1 participant
Extracranial healthy	6	4 participants	1 participant	1 participant
Neurovascular	3	2 participants	Patch-level ∗	1 participant
Neuroinflammatory	3	2 participants	Patch-level ∗	1 participant

Table S1 : Training, validation, and test splits for in vivo transfer learning.

Figure S2 : Voxel-wise linear regression between high-resolution (HR) reference velocities ( $V_{\mathrm{HR}}$ ) and super-resolved velocities ( $V_{\mathrm{SR}}$ ) at peak systole for 4DFlowNet Ferdian et al. (2023) , the CFD-pretrained baseline (CFD), the cohort-specific fine-tuned model (COHORT), and the pooled fine-tuned model (POOL) across the four test cases. For the intracranial, neurovascular, and neuroinflammatory cases, signed velocity components ( $V_{x},V_{y},V_{z}$ ) were pooled; for the extracranial case, only $V_{z}$ was analyzed, representing the predominant velocity component. Solid lines indicate fitted linear regressions and dashed lines the identity line. Regression metrics were calculated using all analyzed voxels within the vessel mask. For visualization, an equal number of voxels per artery was sampled. Points are colored according to artery label.

Medical Image Super-Resolution

Medical image super-resolution reconstructs fine spatial detail from low-resolution scans to support clearer anatomy and more useful clinical measurements.

See all 57 papers

Distributional Deep Learning for Super-Resolution of 4D Flow MRI under Domain Shift
Feb 16, 2026Xiaoyi Wen, Fei JiangMedical Image Super-Resolution Magnetic Resonance Imaging
Department of Epidemiology and Biostatistics, University of California, San Francisco
Attention-guided super-resolution of 4D flow MRI in carotid arteries
Sep 4, 2026Ali Mokhtari, Dominik ObristMedical Image Super-Resolution Magnetic Resonance Imaging
ARTORG Center for Biomedical Engineering Research, University of Bern, Bern, Switzerland
Spectrally optimized implicit neural representations mitigate acquisition-related image quality trade-offs in intracranial 4D Flow MRI
Oct 23, 2026Oliver Welin Odeback, Nerea González-Aranceta, Sebastián Jofre +12Medical Image Super-Resolution Magnetic Resonance Imaging
Dept. Molecular Medicine and Surgery, Karolinska Institutet, Karolinska Universitetssjukhuset Solna (L1:00), Stockholm, 171 76, Sweden. · Department of Physics, University of Greifswald, Felix-Hausdorff-Str. 6, Greifswald, 174 89, Germany. · Biomedical Imaging Center, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, Macul, Santiago, 7820436, Chile. +7

Feb 16, 2026J/K move · Enter open · S save

Distributional Deep Learning for Super-Resolution of 4D Flow MRI under Domain Shift

Xiaoyi Wen, Fei Jiang

Department of Epidemiology and Biostatistics, University of California, San Francisco

Medical Image Super-Resolution Magnetic Resonance Imaging Domain Generalization

Open paper AskPDF arXiv

Super-resolution is widely used in medical imaging to enhance low-quality data, reducing scan time and improving abnormality detection. Conventional super-resolution approaches typically rely on paired datasets of downsampled and original high resolution images, training models to reconstruct high resolution images from their artificially degraded counterparts. However, in real-world clinical settings, low resolution data often arise from acquisition mechanisms that differ significantly from simple downsampling. As a result, these inputs may lie outside the domain of the training data, leading to poor model generalization due to domain shift. To address this limitation, we propose a distributional deep learning framework that improves model robustness and domain generalization. We develop this approch for enhancing the resolution of 4D Flow MRI (4DF). This is a novel imaging modality that captures hemodynamic flow velocity and clinically relevant metrics such as vessel wall stress. These metrics are critical for assessing aneurysm rupture risk. Our model is initially trained on high resolution computational fluid dynamics (CFD) simulations and their downsampled counterparts. It is then fine-tuned on a small, harmonized dataset of paired 4D Flow MRI and CFD samples. We derive the theoretical properties of our distributional estimators and demonstrate that our framework significantly outperforms traditional deep learning approaches through real data applications. This highlights the effectiveness of distributional learning in addressing domain shift and improving super-resolution performance in clinically realistic scenarios.

Abstract

Medical Image Super-Resolution

Explore similar work

Distributional Deep Learning for Super-Resolution of 4D Flow MRI under Domain Shift

Attention-guided super-resolution of 4D flow MRI in carotid arteries

Spectrally optimized implicit neural representations mitigate acquisition-related image quality trade-offs in intracranial 4D Flow MRI