Revolutionizing Brain Tumor Imaging: Generating Synthetic 3D FA Maps from T1-Weighted MRI using CycleGAN Models

Francesca M. Cozzi; Rajesh Jena; Xin Du

REVIEW 6 major objections 6 minor 33 references

Revolutionizing Brain Tumor Imaging: Generating Synthetic 3D FA Maps from T1-Weighted MRI using CycleGAN Models

T0 review · 6 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a 3D CycleGAN can synthesize fractional anisotropy (FA) maps from T1-weighted MRI with SSIM around 0.88 in both healthy and tumor brains.

desk verdict A plausible 3D CycleGAN application to FA synthesis, but the tumour-region quantitative claim is uncalibrated and the manuscript is missing key dataset and evaluation details. read the letter →

arxiv 2505.03662 v1 pith:UH6BV2JG submitted 2025-05-06 cs.CV cs.AI

Xin Du , Francesca M. Cozzi , Rajesh Jena This is my paper · ORCID

classification cs.CVcs.AI

keywords CycleGANfractionalanisotropyT1-weightedMRIimage-to-imagetranslationbraintumorwhitematterintegrity3Dsynthesissynthetic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a 3D CycleGAN can synthesize fractional anisotropy (FA) maps, a diffusion-MRI measure of white matter integrity, directly from routine T1-weighted MRI, for both healthy brains and brains with tumors. The authors report SSIM of 0.879 (healthy FA) and 0.883 (tumor FA) against real diffusion-derived maps, with slightly higher values when transfer learning from healthy to tumor cases is used. If valid, this would let clinicians obtain a diffusion-derived biomarker without acquiring a diffusion scan, easing spatial alignment to atlases and reducing patient burden. The paper also shows the synthetic maps contain enough tumor-related structure for a U-Net trained on real FA maps to segment tumor subregions, though with lower Dice than on real maps.

What carries the argument

The load-bearing mechanism is a 3D CycleGAN with two generators and two discriminators trained on unpaired T1 and FA volumes, using three losses: adversarial loss (least-squares), cycle-consistency loss (L1 between round-trip translations), and a correlation-coefficient loss that maximizes linear correlation between generated and real images. The cycle-consistency term forces the mapping to preserve anatomical content, while the correlation-coefficient loss (Eq. 3) is what the authors credit for capturing structural similarity beyond pixel-wise error. The same architecture is applied to DEC maps with three output channels.

What would settle it

Measure voxel-wise error of the synthetic FA maps specifically within peritumoral edema, where diffusion anisotropy is known to drop; if the synthetic maps do not reproduce that regional anisotropy decrease, the claim that the model captures tumor-related microstructure fails.

Watch

Extended reading notes

Core claim

The central claim is that an unpaired image-to-image translation model based on CycleGAN can generate three-dimensional FA and DEC maps from T1-weighted MRI that are quantitatively similar to maps derived from diffusion tensor imaging, for both healthy subjects and brain-tumor patients. The authors support this with voxel-wise intensity comparisons, SSIM/MS-SSIM/PSNR scores, and a segmentation-transfer experiment; they also report that transfer learning from healthy to tumor cases improves tumor-region fidelity and that the model captures tumor structure better than surrounding edited tissue. The paper positions this as the first 3D application of CycleGAN to synthesize FA and DEC maps in both healthy and tumor-affected brains.

Load-bearing premise

The method assumes that T1-weighted anatomical contrast carries enough microstructural information to determine fractional anisotropy, especially inside or near brain tumors where edema and infiltration can decouple the T1 signal from white-matter tract integrity.

Editorial extensions

If this is right

If the paper is right, a T1-only imaging protocol could provide FA and DEC maps for tractography-based models, removing the need for separate diffusion acquisition in many routine studies.
The reported SSIM of about 0.88 on tumor FA means the synthetic maps may be adequate for atlas-based spatial normalization and for feeding downstream predictive models, though local tract detail is still imperfect.
Transfer learning from healthy to tumor cases yields a small but consistent improvement (transfer tumor SSIM 0.886 vs 0.883), suggesting healthy pretraining can help when tumor training data are scarce.
The segmentation-transfer result implies synthetic FA maps preserve enough pathological contrast to be used as input to tumor-segmentation networks, at least for edema (Dice 0.42) though less for enhancing and necrotic regions.
The lower scores for DEC maps (SSIM 0.700) indicate that directionally encoded colour synthesis is harder than scalar FA synthesis, so DEC generation is not yet at the same fidelity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension the paper leaves implicit: replacing the CycleGAN backbone with a conditional diffusion model may close the gap between healthy and tumor SSIM, since diffusion models handle uncertainty in tumor infiltration better than adversarial-only training.
The segmentation Dice for peritumoral edema (0.42) suggests that FA maps carry information about edema that T1 alone might not show explicitly; if replicated, this could hint that synthetic FA provides a surrogate for FLAIR contrast in tumor segmentation.
Because the method trains on unpaired data, it opens the possibility of domain adaptation across scanners and field strengths without paired registration, a natural next step the authors do not mention.
The paper's reliance on SSIM and PSNR may understate errors in small fiber bundles; a future test could compute tract-profile correlations along specific white-matter tracts (e.g., corticospinal tract) from the synthetic FA maps.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a straightforward 3D CycleGAN application for synthesizing FA and DEC maps from T1-weighted MRI, extended to tumour-affected brains, with a transfer-learning comparison and a downstream segmentation test. The core idea is not new (2D FA-from-T1 exists in Gu et al.), and the method is standard CycleGAN plus Cor-Coe loss. The honest value is the 3D, both-healthy-and-tumour scope and the attempt to show clinical utility. That said, the central quantitative claim—'particularly robust performance in tumour regions'—does not hold up as reported. The stress-test concern is real: SSIM and PSNR in a small, low-variance tumour ROI can be high even for a trivial constant prediction. Figure 5 shows tumour-region metrics higher than whole-brain, exactly the artifact signature, but there are no ROI volumes, no FA mean/variance for real vs synthetic, and no baseline comparison. Without a constant-value or paired-data baseline, the claim is unsupported. This is not a fatal flaw in the method; it is a missing-control problem in the evaluation.

Other soft spots are in proportion. The tumour dataset is never described: no source, number of cases, preprocessing, or registration. The abstract says 'rigorously evaluated using SSIM and PSNR,' but PSNR is only shown in figures, not numerically, and the radiologist assessment promised in Section 4.7 never appears. The segmentation downstream test is a nice thought but the Dice is low (0.28 synthetic vs 0.57 real) and the authors themselves call it an initial assessment. On the credit side, the healthy-case experimental setup is clear (HCP, 1000/65 split), and the loss functions and training details are standard and reproducible. The Bland-Altman and transfer-learning analyses are reasonable.

Who is this for? A clinical imaging audience interested in whether routine T1 can substitute for diffusion sequences. The paper would benefit that reader if the evaluation gaps were filled. As it stands, I'd send it to peer review—it deserves referees' time because the application is plausible and the method is sound—but I'd expect major revision: describe the tumour cohort, add trivial and paired baselines, report PSNR numerically, and either include the radiologist data or drop the claim. I would not cite it in its current form.

Referee Report

6 major / 6 minor

Summary. This manuscript proposes a 3D CycleGAN for generating fractional anisotropy (FA) and directionally encoded colour (DEC) maps from T1-weighted MRI, for both healthy and tumor-affected brains. The method uses standard CycleGAN components: two generators and discriminators, an adversarial loss, a cycle-consistency loss, and a correlation-coefficient loss. Healthy data come from 1,065 HCP subjects, split into 1,000 training and 65 test cases; tumor data are mentioned but not described. Quantitative results are reported as SSIM/MS-SSIM in Table 1 (healthy FA 0.879/0.975, tumor FA 0.883/0.958, transfer tumor FA 0.886/0.964, healthy DEC 0.700/0.932). PSNR appears only in figures, and a radiologist visual assessment is promised in Section 4.7 but never reported. A U-Net segmentation transfer experiment is also described. The central claim is that the model produces high-fidelity maps that have been rigorously evaluated, with particularly robust performance in tumour regions.

Significance. If the central claim were fully supported, this would be a useful contribution: it appears to be the first 3D CycleGAN applied to FA/DEC synthesis from T1WI for both healthy and tumor cases, and the idea of generating spatially aligned diffusion-derived maps from routine T1 scans has practical clinical value. Strengths include training on unpaired data with standard losses, evaluation against independently fitted ground-truth FA maps (no apparent circularity), and a downstream segmentation transfer test. However, the current evidence does not substantiate the headline claims of rigorous evaluation and tumor-region robustness. The missing tumor dataset description, absent baseline comparisons, unreported numerical PSNR values, and unreported radiologist results are substantial gaps that prevent the paper from supporting its abstract.

major comments (6)

[Section 3 / Table 1 / Figure 5] The tumor-affected dataset is never defined. The paper describes only the healthy HCP dataset; no information is given on the source, number of tumor subjects, tumor types, acquisition parameters, preprocessing, or how T1WI and ground-truth FA maps were registered. Consequently, the tumor FA rows in Table 1 (SSIM 0.883 and 0.886) and the tumor-region analysis in Figure 5 cannot be reproduced or evaluated.
[Abstract / Section 5.1 / Figure 5] The claim that performance is 'particularly robust in tumour regions' is not supported by the data shown. In low-variance regions such as tumor core and edema, SSIM and PSNR are inflated even for a trivial constant predictor because these metrics are dominated by intensity range and variance. The paper reports no ROI volumes, no mean and standard deviation of FA in real versus synthetic tumor regions, no trivial baseline, and no paired-data baseline; the observation that tumor-region metrics exceed whole-brain metrics is exactly the expected signature of this artifact.
[Section 5.1 / Table 1] The narrative is internally inconsistent. The text states that the model performs better for healthy brains and that the difference between generated and ground-truth maps is 'statistically lower in healthy cases', yet Table 1 reports comparable or higher SSIM for tumor FA (0.883 vs. 0.879) and the abstract claims particularly robust tumor performance. The authors should state a single primary claim and report the statistics that adjudicate between these statements.
[Section 4.8 / Figure 3] PSNR is central to the abstract's claim of 'rigorously evaluated using SSIM and PSNR', but the paper reports no numerical PSNR values, no standard deviations, and no number of test cases; PSNR appears only in figures. The authors should report mean and standard deviation of PSNR per condition, with case counts and, where appropriate, paired statistical tests, in a table alongside the SSIM/MS-SSIM results.
[Section 5 / no baseline comparisons] The paper reports no comparison against existing methods such as the 2D CycleGAN of Gu et al. [19], a supervised U-Net, or a simple regression from T1 intensity. Without a baseline, the favorable claims about the 3D CycleGAN's ability to synthesize FA maps are not established; at minimum, a comparison with the prior 2D method is needed to justify the '3D' contribution.
[Section 4.7 / Section 5] Section 4.7 promises a visual assessment by two radiologists distinguishing generated from ground-truth maps, and the abstract states that 'radiological assessments further underscore' the model's potential, but Section 5 reports no radiologist results. The authors should either include the full protocol and outcomes (blinding, sample size, inter-rater agreement, and quantitative results) or remove these claims from the abstract and methods.

minor comments (6)

[Section 4.7 / 4.8] Section 4.8 is numbered as a subsection of 'Evaluation Metrics' but appears after Section 4.7.3; renumber it as 4.7.3 or a separate top-level section to avoid confusion.
[Table 1] The labels 'Health FA' and 'Tumor FA' should be 'Healthy FA' and 'Tumour FA' for consistency with the rest of the text.
[Equation (3)] Equation (3) uses covariance and standard deviations without defining whether these are computed over the whole volume or over a local window; define the computation clearly.
[Equation (4) / Section 4.6] The hyperparameters λ and β in Eq. (4) are both reported as 1 with no sensitivity analysis or ablation; a brief ablation or a reference justifying these values would strengthen the results.
[Figure 2] The Bland-Altman plot appears to compare mean intensities per scan rather than voxel-wise differences; clarify the unit of analysis and the exact quantity plotted.
[Introduction / Table 1] The introduction promises DEC maps as a central output, but the DEC evaluation is limited to one row of Table 1; the text should state whether DEC generation is a primary or secondary outcome.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported predictions are evaluated against external ground-truth FA maps using metrics distinct from the training objective.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. The CycleGAN is trained with a standard objective (Equation 4) combining adversarial loss, cycle-consistency loss, and a correlation-coefficient loss, with weights lambda = beta = 1. The training objective is not the evaluation metric: evaluation uses SSIM, MS-SSIM, and PSNR (Equations 5, 6, and 7), which are external to the loss function. No model parameter is fitted to the reported SSIM/PSNR values, and the healthy test set is explicitly held out (65 unseen HCP subjects, Section 3). Ground-truth FA maps are derived from diffusion tensor and MAP-MRI fitting following Gu et al. [19], an external prior work, not from the authors' own models. The tumor segmentation experiment even provides an independent transfer test: a U-Net trained on real FA maps is applied to synthetic FA maps, with the lower Dice scores reported rather than used as a training target. There is no self-citation chain: references [11], [19], [28], and [29] are prior works by other authors, and no uniqueness theorem or ansatz is imported from the current authors' earlier publications. The manuscript does have significant reporting gaps—the tumor dataset is not described in Section 3, the promised radiologist assessment in Section 4.7 is absent, and the claim of 'particularly robust performance in tumour regions' is not calibrated against a trivial baseline. These are correctness and evidence concerns, not circularity, because they do not make any derived quantity equivalent to an input by construction. The central results are externally grounded and the evaluation metrics are not constructed from model outputs, so no circular step is present. Score 0 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the learnability of T1-to-FA mapping, the fidelity of the reference FA maps, and the transferability from healthy to tumor data. These are domain assumptions supported by prior work, not by new independent evidence. No new physical or mathematical entities are introduced.

free parameters (5)

Cycle consistency loss weight lambda = 1
Set to 1 in Section 4.5 without tuning analysis.
Correlation coefficient loss weight beta = 1
Set to 1 in Section 4.5 without tuning analysis.
Learning rate = 0.0002
Adam optimizer learning rate chosen in Section 4.6.
Batch size = 1 or 8
Section 4.6 says batch size of 1 or 8 depending on memory constraints; ambiguous.
Input volume size = 128x128x64
All inputs resized to this volume in Section 4.6.

assumptions (4)

domain assumption T1-weighted MRI contains sufficient information to recover fractional anisotropy.
The entire mapping T1 to FA assumes this; Section 1 calls T1WI an 'ideal candidate' without evidence.
domain assumption Ground-truth FA maps obtained by diffusion tensor and MAP-MRI fitting are accurate references.
Section 3 states ground truth is obtained via fitting as in Gu et al.; evaluation quality depends on this reference.
domain assumption A model trained on healthy HCP subjects can transfer to tumor-affected brains.
Transfer learning from healthy to tumor is used in Section 4.6 and evaluated in Section 5; domain shift is assumed manageable.
domain assumption SSIM, MS-SSIM, and PSNR are adequate proxies for clinical usefulness of synthetic FA maps.
Section 4.7 uses these metrics as the main quantitative evidence; no clinical task validation is performed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Revolutionizing Brain Tumor Imaging: Generating Synthetic 3D FA Maps from T1-Weighted MRI using CycleGAN Models." pith.science (2026). https://pith.science/paper/UH6BV2JG

@misc{pith2026250503662,
  author       = {Pith},
  title        = {Pith review of: Revolutionizing Brain Tumor Imaging: Generating Synthetic 3D FA Maps from T1-Weighted MRI using CycleGAN Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UH6BV2JG}},
  note         = {Machine review of arXiv:2505.03662}
}

read the original abstract

Fractional anisotropy (FA) and directionally encoded colour (DEC) maps are essential for evaluating white matter integrity and structural connectivity in neuroimaging. However, the spatial misalignment between FA maps and tractography atlases hinders their effective integration into predictive models. To address this issue, we propose a CycleGAN based approach for generating FA maps directly from T1-weighted MRI scans, representing the first application of this technique to both healthy and tumour-affected tissues. Our model, trained on unpaired data, produces high fidelity maps, which have been rigorously evaluated using Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR), demonstrating particularly robust performance in tumour regions. Radiological assessments further underscore the model's potential to enhance clinical workflows by providing an AI-driven alternative that reduces the necessity for additional scans.

Figures

Figures reproduced from arXiv: 2505.03662 by the authors.

**Figure 3.** Quantitative evaluation of the generated [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 2.** Bland-Altman plot comparing the generated FA maps with the ground truth for both healthy and tumor cases (encompassing results of learning from scratch and transfer learning approaches) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Comparison of generated images with and without tumour masking based on PSNR evaluation. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Comparison of generated image quality between the tumour region, the region after tumour [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Illustration of tumour segmentation. performance is not the primary focus of this work. 6 Conclusion In this study, we investigated the feasibility of generating diffusion tensor imaging (DTI) scans, including fractional anisotropy (FA) and directionally encoded col…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 32 canonical work pages

[19]

X. Gu, H. Knutsson, M. Nilsson, A. Ek- lund, Generating diffusion mri scalar maps from t1 weighted images using generative adversarial networks, in: Image Analysis: 21st Scandinavian Conference, SCIA 2019, Norrk¨oping, Sweden, June 11–13, 2019, Pro- ceedings 21, Springer, 2019, pp. 489–498

work page 2019
[1]

Wang, K.-H

L.-W. Wang, K.-H. Cho, P.-Y . Chao, L.-W. Kuo, C.-W. Chiang, C.-M. Chao, M.-T. Lin, C.-P. Chang, H.-J. Lin, C.-C. Chio, White and gray matter integrity evaluated by mri-dti can serve as noninvasive and reliable indica- tors of structural and functional alterations in chronic neurotrauma, Scientific Reports 14 (2024) 7244

work page 2024
[2]

Y . Feng, B. Q. Chandio, J. E. Villalon- Reina, S. I. Thomopoulos, T. M. Nir, S. Benavidez, E. Laltoo, T. Chattopad- hyay, H. Joshi, G. Venkatasubramanian, et al., Microstructural mapping of neu- ral pathways in alzheimer’s disease using macrostructure-informed normative tractom- etry, Alzheimer’s & Dementia 21 (2025) e14371

work page 2025
[3]

P. J. Basser, S. Pajevic, C. Pierpaoli, J. Duda, A. Aldroubi, In vivo fiber tractography using dt-mri data, Magnetic resonance in medicine 44 (2000) 625–632

work page 2000
[4]

Le Bihan, M

D. Le Bihan, M. Iima, Diffusion magnetic resonance imaging: what water tells us about biological tissues, PLoS biology 13 (2015) e1002203

work page 2015
[5]

D. C. Van Essen, D. L. Dierker, Surface- based and probabilistic atlases of primate cerebral cortex, Neuron 56 (2007) 209–225

work page 2007
[6]

D. K. Jones, T. R. Kn ¨osche, R. Turner, White matter integrity, fiber count, and other fal- lacies: the do’s and don’ts of diffusion mri, Neuroimage 73 (2013) 239–254

work page 2013
[7]

Karimi, S

D. Karimi, S. K. Warfield, Diffusion mri with machine learning, Imaging Neuroscience 2 (2024) 1–55

work page 2024

Show all 33 references

[8]

C. J. Preetha, H. Meredig, G. Brugnara, M. A. Mahmutoglu, M. Foltyn, F. Isensee, T. Kessler, I. Pfl ¨uger, M. Schell, U. Neu- berger, et al., Deep-learning-based synthesis of post-contrast t1-weighted mri for tumour response assessment in neuro-oncology: a multicentre, retrosp...

2021
[9]

Herrmann, T

J. Herrmann, T. Benkert, A. Brendlin, S. Gassenmaier, T. H¨olldobler, S. Maennlin, H. Almansour, A. Lingg, E. Weiland, S. Afat, Shortening acquisition time and improv- ing image quality for pelvic mri using deep learning reconstruction for diffusion- weighted imaging at 1.5 t,...

2024
[10]

Abu-Srhan, I

A. Abu-Srhan, I. Almallahi, M. A. Abushariah, W. Mahafza, O. S. Al-Kadi, Paired-unpaired unsupervised attention guided gan with transfer learning for bidirec- tional brain mr-ct synthesis, Computers in Biology and Medicine 136 (2021) 104763

2021
[11]

J.-Y . Zhu, T. Park, P. Isola, A. A. Efros, Unpaired image-to-image translation using cycle-consistent adversarial networks, in: Proceedings of the IEEE international con- ference on computer vision, 2017, pp. 2223– 2232

2017
[12]

Dayarathna, K

S. Dayarathna, K. T. Islam, S. Uribe, G. Yang, M. Hayat, Z. Chen, Deep learn- ing based synthesis of mri, ct and pet: Re- view and analysis, Medical image analysis 92 (2024) 103046

2024
[13]

Y . Hu, H. Zhou, N. Cao, C. Li, C. Hu, Syn- thetic ct generation based on cbct using im- proved vision transformer cyclegan, Scien- tific Reports 14 (2024) 11455

2024
[14]

Diniz, T

E. Diniz, T. Santini, K. Helmet, H. J. Aizen- stein, T. S. Ibrahim, Cross-modality im- age translation of 3 tesla magnetic resonance imaging to 7 tesla using generative adversar- ial networks, medRxiv (2024)

2024
[15]

C. Gong, Y . Huang, M. Luo, S. Cao, X. Gong, S. Ding, X. Yuan, W. Zheng, Y . Zhang, Channel-wise attention enhanced and structural similarity constrained cycle- gan for effective synthetic ct generation from head and neck mri images, Radiation Oncol- ogy 19 (2024) 37

2024
[16]

J. M. Wolterink, A. M. Dinkla, M. H. Savenije, P. R. Seevinck, C. A. van den Berg, I. I ˇsgum, Deep MR to ct synthesis using unpaired data, in: Simulation and Synthe- sis in Medical Imaging: Second International Workshop, SASHIMI 2017, Held in Con- junction with MICCAI 2017, Q...

2017
[17]

J. Wang, Q. J. Wu, F. Pourpanah, Dc- cyclegan: Bidirectional ct-to-mr synthesis from unpaired data, Computerized Medical Imaging and Graphics 108 (2023) 102249

2023
[18]

A. H. Farshchitabrizi, M. H. Sadeghi, S. Sina, M. Alavi, Z. N. Feshani, H. Omidi, Ai-enhanced pet/ct image synthesis using cy- clegan for improved ovarian cancer imaging, Polish Journal of Radiology 90 (2025) 26– 35

2025
[20]

Huang, S

X. Huang, S. Belongie, Arbitrary style trans- fer in real-time with adaptive instance nor- malization, in: Proceedings of the IEEE in- ternational conference on computer vision, 2017, pp. 1501–1510

2017
[21]

S. Cao, N. Konz, J. Duncan, M. A. Mazurowski, Deep learning for breast mri style transfer with limited training data, Jour- nal of Digital imaging 36 (2023) 666–678

2023
[22]

Kazerouni, E

A. Kazerouni, E. K. Aghdam, M. Hei- dari, R. Azad, M. Fayyaz, I. Hacihaliloglu, D. Merhof, Diffusion models in medical imaging: A comprehensive survey, Medical image analysis 88 (2023) 102846

2023
[23]

Dhariwal, A

P. Dhariwal, A. Nichol, Diffusion models beat gans on image synthesis, Advances in neural information processing systems 34 (2021) 8780–8794

2021
[24]

J. Kim, H. Park, Adaptive latent diffusion model for 3d medical image to image transla- tion: Multi-modal magnetic resonance imag- ing study, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision, 2024, pp. 7604–7613

2024
[25]

M. F. Glasser, S. N. Sotiropoulos, J. A. Wil- son, T. S. Coalson, B. Fischl, J. L. Ander- sson, J. Xu, S. Jbabdi, M. Webster, J. R. Polimeni, et al., The minimal preprocessing pipelines for the human connectome project, Neuroimage 80 (2013) 105–124

2013
[26]

D. C. Van Essen, S. M. Smith, D. M. Barch, T. E. Behrens, E. Yacoub, K. Ugurbil, W.- M. H. Consortium, et al., The wu-minn hu- man connectome project: an overview, Neu- roimage 80 (2013) 62–79

2013
[27]

Y . Ge, D. Wei, Z. Xue, Q. Wang, X. Zhou, Y . Zhan, S. Liao, Unpaired MR to CT syn- thesis with explicit structural constrained ad- versarial learning, in: 2019 IEEE 16th Inter- national Symposium on Biomedical Imaging (ISBI 2019), IEEE, 2019, pp. 1096–1099

2019
[28]

Y . Ge, Z. Xue, T. Cao, S. Liao, Unpaired whole-body mr to ct synthesis with correla- tion coefficient constrained adversarial learn- ing, in: Medical Imaging 2019: Image Pro- cessing, volume 10949, SPIE, 2019, pp. 28– 35

2019
[29]

Johnson, A

J. Johnson, A. Alahi, L. Fei-Fei, Perceptual losses for real-time style transfer and super- resolution, in: Computer Vision–ECCV 2016: 14th European Conference, Amster- dam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, Springer, 2016, pp. 694–711

2016
[30]

Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE transactions on image processing 13 (2004) 600–612

2004
[31]

Z. Wang, E. P. Simoncelli, A. C. Bovik, Mul- tiscale structural similarity for image qual- ity assessment, in: The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, Ieee, 2003, pp. 1398–1402

2003
[32]

A. Hore, D. Ziou, Image quality met- rics: Psnr vs. ssim, in: 2010 20th inter- national conference on pattern recognition, IEEE, 2010, pp. 2366–2369

2010
[33]

Isensee, P

F. Isensee, P. F. Jaeger, S. A. Kohl, J. Pe- tersen, K. H. Maier-Hein, nnu-net: a self- configuring method for deep learning-based biomedical image segmentation, Nature methods 18 (2021) 203–211

2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [19]

X. Gu, H. Knutsson, M. Nilsson, A. Ek- lund, Generating diffusion mri scalar maps from t1 weighted images using generative adversarial networks, in: Image Analysis: 21st Scandinavian Conference, SCIA 2019, Norrk¨oping, Sweden, June 11–13, 2019, Pro- ceedings 21, Springer, 2019, pp. 489–498

work page 2019

[2] [1]

Wang, K.-H

L.-W. Wang, K.-H. Cho, P.-Y . Chao, L.-W. Kuo, C.-W. Chiang, C.-M. Chao, M.-T. Lin, C.-P. Chang, H.-J. Lin, C.-C. Chio, White and gray matter integrity evaluated by mri-dti can serve as noninvasive and reliable indica- tors of structural and functional alterations in chronic neurotrauma, Scientific Reports 14 (2024) 7244

work page 2024

[3] [2]

Y . Feng, B. Q. Chandio, J. E. Villalon- Reina, S. I. Thomopoulos, T. M. Nir, S. Benavidez, E. Laltoo, T. Chattopad- hyay, H. Joshi, G. Venkatasubramanian, et al., Microstructural mapping of neu- ral pathways in alzheimer’s disease using macrostructure-informed normative tractom- etry, Alzheimer’s & Dementia 21 (2025) e14371

work page 2025

[4] [3]

P. J. Basser, S. Pajevic, C. Pierpaoli, J. Duda, A. Aldroubi, In vivo fiber tractography using dt-mri data, Magnetic resonance in medicine 44 (2000) 625–632

work page 2000

[5] [4]

Le Bihan, M

D. Le Bihan, M. Iima, Diffusion magnetic resonance imaging: what water tells us about biological tissues, PLoS biology 13 (2015) e1002203

work page 2015

[6] [5]

D. C. Van Essen, D. L. Dierker, Surface- based and probabilistic atlases of primate cerebral cortex, Neuron 56 (2007) 209–225

work page 2007

[7] [6]

D. K. Jones, T. R. Kn ¨osche, R. Turner, White matter integrity, fiber count, and other fal- lacies: the do’s and don’ts of diffusion mri, Neuroimage 73 (2013) 239–254

work page 2013

[8] [7]

Karimi, S

D. Karimi, S. K. Warfield, Diffusion mri with machine learning, Imaging Neuroscience 2 (2024) 1–55

work page 2024