arxiv: 2604.14800 · v1 · submitted 2026-04-16 · 📡 eess.IV · cs.CV· physics.med-ph

Recognition: unknown

Generative Modeling of Complex-Valued Brain MRI Data

Jens Kleesiek, Jens Weingarten, Jessica Mnischek, Kevin Kr\"oninger, Lukas T. Rotkopf, Marco Schlimbach, Moritz Rempe

Authors on Pith no claims yet

Pith reviewed 2026-05-10 09:18 UTC · model grok-4.3

classification 📡 eess.IV cs.CVphysics.med-ph

keywords complex-valued MRIgenerative modelingphase preservationsynthetic data augmentationbrain abnormality detectionvariational autoencoderflow matchingfastMRI dataset

0 comments

The pith

A generative model for complex-valued brain MRI produces synthetic scans that let classifiers detect abnormalities more accurately than training on real data alone.

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

The paper sets out to show that magnitude and phase can be modeled together in a generative system for brain MRI instead of throwing away phase as standard pipelines do. It pairs a conditional variational autoencoder that keeps phase coherence above 0.997 with a flow-matching generator to create new samples. Downstream classifiers trained only on these synthetic images reach an AUROC of 0.880 for normal-versus-abnormal detection on fastMRI, beating the 0.842 score from real data, and the edge remains on an external biopsy-labeled set. A reader would care because real MRI datasets are small and costly, while phase is known to carry tissue information that could sharpen diagnosis if it can be used reliably.

Core claim

The framework uses a conditional variational autoencoder to compress complex-valued MRI into latent codes while preserving phase coherence above 0.997, then applies a flow-matching model to generate new complex scans. Real-versus-synthetic classifiers give AUROC scores between 0.50 and 0.66, showing the outputs are nearly indistinguishable from real data. Classifiers trained entirely on the synthetic data achieve an AUROC of 0.880 for normal-versus-abnormal classification on the fastMRI dataset, surpassing the 0.842 baseline obtained from real data, and this performance advantage carries over to an independent external test set with biopsy-confirmed labels.

What carries the argument

Conditional variational autoencoder that preserves phase coherence in complex MRI, paired with a flow-matching generative model that samples new magnitude-plus-phase images.

If this is right

Synthetic complex-valued MRI can augment scarce real datasets for training diagnostic models without loss of performance.
Phase data in MRI contains features useful for distinguishing normal from abnormal tissue that magnitude images alone miss.
The same generative approach can supply labeled training examples for rare conditions where collecting real scans is difficult.
Generated samples maintain utility across different scanners and institutions, supporting multi-site model development.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If phase truly improves detection, reconstruction software could be redesigned to output and store complex data rather than magnitude only.
The same joint magnitude-phase modeling could be tested on other modalities that acquire complex signals, such as certain CT or ultrasound techniques.
A controlled ablation that removes phase from the synthetic data and measures the resulting drop in classifier AUROC would isolate how much of the gain comes from phase versus the generative process.

Load-bearing premise

The phase information encoded and regenerated by the model carries genuine pathology signals that improve classification without introducing spurious patterns that happen to work only on the chosen test sets.

What would settle it

Retraining the downstream normal-versus-abnormal classifier on a larger, multi-institution real dataset and finding that its AUROC exceeds the synthetic-data AUROC of 0.880 would show the reported advantage does not generalize.

Figures

Figures reproduced from arXiv: 2604.14800 by Jens Kleesiek, Jens Weingarten, Jessica Mnischek, Kevin Kr\"oninger, Lukas T. Rotkopf, Marco Schlimbach, Moritz Rempe.

**Figure 2.** Figure 2: Two samples of training data after ESPIRiT combination, IFT and patching for a normal [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Autoencoder reconstruction examples for an abnormal sample (AXT1POST, left) and a [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Generated and real samples from the Stage 1 flow matching model. Synthetic samples [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: Samples generated by the Stage 2 flow matching model, conditioned on normal (left) and [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Downstream classification AUROC on the fastMRI test set as a function of real data [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

**Figure 7.** Figure 7: Downstream classification AUROC on the external test set as a function of real data [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

**Figure 8.** Figure 8: Downstream classification AUROC when synthetic data is progressively added to the full [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

read the original abstract

Objective. Standard Magnetic Resonance Imaging (MRI) reconstruction pipelines discard phase information captured during acquisition, despite evidence that it encodes tissue properties relevant to tumor diagnosis. Current machine learning approaches inherit this limitation by operating exclusively on reconstructed magnitude images. The aim of this study is to build a generative framework which is capable of jointly modeling magnitude and phase information of complex-valued MRI scans. Approach. The proposed generative framework combines a conditional variational autoencoder, which compresses complex-valued MRI scans into compact latent representations while preserving phase coherence, with a flow-matching-based generative model. Synthetic sample quality is assessed via a real-versus-synthetic classifier and by training downstream classifiers on synthetic data for abnormal tissue detection. Main results. The autoencoder preserves phase coherence above 0.997. Real-versus-synthetic classification yields low AUROC values between 0.50 and 0.66 across all acquisition sequences, indicating generated samples are nearly indistinguishable from real data. In downstream normal-versus-abnormal classification, classifiers trained entirely on synthetic data achieve an AUROC of 0.880, surpassing the real-data baseline of 0.842 on a publicly available dataset (fastMRI). This advantage persists on an independent external test set from a different institution with biopsy-confirmed labels. Significance. The proposed framework demonstrates the feasibility of jointly modeling magnitude and phase information for normal and abnormal complex-valued brain MRI data. Beyond synthetic data generation, it establishes a foundation for the usage of complete brain MRI information in future diagnostic applications and enables systematic investigation of how magnitude and phase jointly encode pathology-specific features.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gets complex-valued MRI generation working with a cVAE-plus-flow setup and reports synthetic data beating real fastMRI baselines downstream, but the phase contribution still needs direct checks.

read the letter

The main takeaway is that they built a conditional VAE to compress complex brain MRI while keeping phase coherence above 0.997, then used flow-matching to generate new samples. Real-versus-synthetic classifiers struggle (AUROC 0.50-0.66), and classifiers trained only on the synthetic complex data reach 0.880 AUROC on normal-vs-abnormal detection versus 0.842 on real fastMRI data, with the edge holding on an external biopsy-labeled set. That downstream lift is the part worth paying attention to if it survives scrutiny. What works is the joint modeling of magnitude and phase in one pipeline; prior work mostly stayed on magnitude images, so this is a concrete step toward using the full signal. The numbers on sample realism and the external-set result are the strongest evidence they provide. The soft spots are the lack of a magnitude-only ablation to show the phase is actually driving the gain, plus limited detail on how the conditioning variables and network capacity were chosen. Without those, it is hard to rule out that the improvement comes from better overall data augmentation rather than pathology-relevant phase structure. The external test helps, but distribution shift between training and that scanner still needs more attention. This is the kind of paper that belongs in a reading group focused on medical imaging or generative models for MRI. It is worth a serious referee because the core idea is new enough and the empirical claims are testable, even if the current write-up leaves some controls for later rounds. I would send it out rather than desk-reject.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes a generative framework that combines a conditional variational autoencoder (to compress complex-valued brain MRI while preserving phase) with a flow-matching model to synthesize realistic magnitude-and-phase data. It reports phase coherence above 0.997, real-versus-synthetic AUROCs of 0.50–0.66, and downstream normal-versus-abnormal classification AUROCs of 0.880 (synthetic) versus 0.842 (real) on fastMRI, with the advantage persisting on an external biopsy-labeled dataset from a different institution.

Significance. If the central empirical claims hold after verification, the work would be significant for demonstrating that joint modeling of magnitude and phase can produce synthetic data that improves downstream diagnostic performance beyond magnitude-only baselines. The use of a public dataset (fastMRI) and an independent external test set with biopsy confirmation strengthens the evaluation.

major comments (3)

[Downstream evaluation] Downstream evaluation section: the AUROC gain (0.880 synthetic vs. 0.842 real) is presented as evidence that phase information contributes diagnostically relevant features, yet no ablation comparing complex-valued synthetic data against magnitude-only synthetic data is reported. This comparison is required to establish that the observed lift is attributable to phase modeling rather than other aspects of the generative pipeline.
[External validation] External validation paragraph: the persistence of the performance advantage on the independent biopsy-labeled set is a key claim, but the manuscript provides no quantitative assessment of distribution shift (e.g., scanner parameters, acquisition protocols) between the generative training distribution and the external data. Without this, it remains possible that the reported edge reflects unmodeled covariates rather than genuine pathology encoding.
[Methods] Methods, model architecture and training: the description of the conditional VAE latent dimension, conditioning variables, flow-matching network capacity, and training schedule is insufficient to reproduce the reported phase coherence >0.997 or to diagnose whether the downstream gains arise from faithful pathology encoding or from modeling artifacts.

minor comments (1)

[Abstract] The abstract states real-versus-synthetic AUROC values as a range (0.50–0.66) without per-sequence breakdowns or reference to a supplementary table; adding this detail would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects for strengthening our claims on phase-aware generative modeling of complex-valued MRI. We address each major comment below and will revise the manuscript to incorporate the requested analyses and details.

read point-by-point responses

Referee: [Downstream evaluation] Downstream evaluation section: the AUROC gain (0.880 synthetic vs. 0.842 real) is presented as evidence that phase information contributes diagnostically relevant features, yet no ablation comparing complex-valued synthetic data against magnitude-only synthetic data is reported. This comparison is required to establish that the observed lift is attributable to phase modeling rather than other aspects of the generative pipeline.

Authors: We agree that the current evidence would be strengthened by directly isolating the contribution of phase. In the revised manuscript we will add an ablation in which a magnitude-only generative model (trained and sampled without phase) is compared head-to-head with the complex-valued model on the same downstream abnormality-classification task. This will clarify whether the reported AUROC advantage arises from joint magnitude-phase modeling. revision: yes
Referee: [External validation] External validation paragraph: the persistence of the performance advantage on the independent biopsy-labeled set is a key claim, but the manuscript provides no quantitative assessment of distribution shift (e.g., scanner parameters, acquisition protocols) between the generative training distribution and the external data. Without this, it remains possible that the reported edge reflects unmodeled covariates rather than genuine pathology encoding.

Authors: We acknowledge the need for explicit quantification of distribution shift. In the revision we will add a dedicated paragraph reporting scanner field strengths, sequence parameters, and statistical measures of image-level shift (intensity histograms, contrast-to-noise ratios, and Kolmogorov-Smirnov tests) between the fastMRI training distribution and the external biopsy-labeled cohort. This analysis will help substantiate that the observed advantage is attributable to pathology encoding rather than acquisition covariates. revision: yes
Referee: [Methods] Methods, model architecture and training: the description of the conditional VAE latent dimension, conditioning variables, flow-matching network capacity, and training schedule is insufficient to reproduce the reported phase coherence >0.997 or to diagnose whether the downstream gains arise from faithful pathology encoding or from modeling artifacts.

Authors: We appreciate the call for greater reproducibility. The revised methods section will specify the cVAE latent dimension, the complete set of conditioning variables, the exact architecture and parameter count of the flow-matching network, and the full training schedule with all hyperparameters. These additions will enable independent reproduction of the phase-coherence results and allow readers to assess whether downstream gains stem from pathology encoding or modeling artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity detected; results are empirical evaluations on held-out data

full rationale

The paper describes an empirical pipeline (conditional VAE + flow-matching) for generating complex-valued MRI and evaluates it via direct metrics: phase coherence >0.997, real-vs-synthetic AUROC 0.50-0.66, and downstream normal-vs-abnormal AUROC of 0.880 (synthetic-trained) vs 0.842 (real baseline) on fastMRI held-out data plus an external biopsy-labeled set. No mathematical derivation chain exists that reduces these quantities to fitted parameters by construction. No self-definitional steps, fitted-input predictions, or load-bearing self-citations are present in the reported results. The external test set and held-out real data provide independent benchmarks, keeping the evaluation self-contained without circular reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions of variational autoencoders and flow-matching models plus the unstated premise that the training data distribution is representative of real clinical variability. No new physical axioms or invented entities are introduced.

free parameters (2)

latent dimension and conditioning variables
Standard hyperparameters of the cVAE that control compression and phase preservation; their specific values are not reported in the abstract.
flow-matching network capacity and training schedule
Hyperparameters of the generative model whose tuning affects sample quality and downstream AUROC.

axioms (1)

domain assumption The complex-valued MRI signal can be meaningfully compressed while preserving phase coherence for downstream pathology tasks.
Invoked in the autoencoder design and phase-coherence metric; not derived from first principles.

pith-pipeline@v0.9.0 · 5606 in / 1399 out tokens · 32763 ms · 2026-05-10T09:18:31.762975+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages · 1 internal anchor

[1]

Siegel, Isabelle Soerjomataram, and Ahmedin Jemal

Freddie Bray, Mathieu Laversanne, Hyuna Sung, Jacques Ferlay, Rebecca L. Siegel, Isabelle Soerjomataram, and Ahmedin Jemal. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA: A Cancer Journal for Clinicians, 74(3):229–263, 2024. https://doi.org/10.3322/caac.21834

work page doi:10.3322/caac.21834 2022
[2]

Jakola, Thomas Skoglund, and Johan Ljungqvist

Anna Pennlund, Asgeir S. Jakola, Thomas Skoglund, and Johan Ljungqvist. A single-centre study of frame-based stereotactic brain biopsies.British Journal of Neurosurgery, 36(2): 213–216, 2022. https://doi.org/10.1080/02688697.2020.1867704

work page doi:10.1080/02688697.2020.1867704 2022
[3]

Luke Dixon, Gurpreet Kaur Jandu, Jai Sidpra, and Kshitij Mankad. Diagnostic accuracy of qualitative MRI in 550 paediatric brain tumours: evaluating current practice in the computational era.Quantitative Imaging in Medicine and Surgery, 12(1):131–143, 2022. https://doi.org/10.21037/qims-20-1388

work page doi:10.21037/qims-20-1388 2022
[4]

Complications after frame-based stereotactic brain biopsy: a systematic review.Neurosurgical Review, 44(1):301–307, 2021

Maximilien Riche, Aymeric Amelot, Matthieu Peyre, Laurent Capelle, Alexandre Carpentier, and Bertrand Mathon. Complications after frame-based stereotactic brain biopsy: a systematic review.Neurosurgical Review, 44(1):301–307, 2021. https://doi.org/10.1007/s10143-019-01234-w

work page doi:10.1007/s10143-019-01234-w 2021
[5]

Badˇ za and MarkoˇC

Milica M. Badˇ za and MarkoˇC. Barjaktarovi´ c. Classification of brain tumors from MRI images using a convolutional neural network.Applied Sciences, 10(6):1999, 2020. https://doi.org/10.3390/app10061999. 14

work page doi:10.3390/app10061999 1999
[6]

Radiomic features and multilayer perceptron network classifier: A robust MRI classification strategy for distinguishing glioblastoma from primary central nervous system lymphoma

Jihye Yun, Ji Eun Park, Hyunna Lee, Sungwon Ham, Namkug Kim, and Ho Sung Kim. Radiomic features and multilayer perceptron network classifier: A robust MRI classification strategy for distinguishing glioblastoma from primary central nervous system lymphoma. Scientific Reports, 9:5746, 2019. https://doi.org/10.1038/s41598-019-42276-w

work page doi:10.1038/s41598-019-42276-w 2019
[7]

Dunarea de Jos

Heba Mohsen, El-Sayed A. El-Dahshan, El-Sayed M. El-Horbaty, and Abdel-Badeeh M. Salem. Brain tumor type classification based on support vector machine in magnetic resonance images.Annals of the “Dunarea de Jos” University of Galati, Fascicle II, Mathematics, Physics, Theoretical Mechanics, 40(1), 2017

2017
[8]

Niakan Kalhori

Soheila Saeedi, Sorayya Rezayi, Hamidreza Keshavarz, and Sharareh R. Niakan Kalhori. MRI-based brain tumor detection using convolutional deep learning methods and chosen machine learning techniques.BMC Medical Informatics and Decision Making, 23(1):16, 2023. https://doi.org/10.1186/s12911-023-02114-6

work page doi:10.1186/s12911-023-02114-6 2023
[9]

Abdusalomov, Mukhriddin Mukhiddinov, and Taeg Keun Whangbo

Akmalbek B. Abdusalomov, Mukhriddin Mukhiddinov, and Taeg Keun Whangbo. Brain tumor detection based on deep learning approaches and magnetic resonance imaging.Cancers, 15(16):4172, 2023. https://doi.org/10.3390/cancers15164172

work page doi:10.3390/cancers15164172 2023
[10]

Mark Haacke, Yingbiao Xu, Yu-Chung N

E. Mark Haacke, Yingbiao Xu, Yu-Chung N. Cheng, and J¨ urgen R. Reichenbach. Susceptibility weighted imaging (SWI).Magnetic Resonance in Medicine, 52(3):612–618, 2004. https://doi.org/10.1002/mrm.20198

work page doi:10.1002/mrm.20198 2004
[11]

Intratumoral susceptibility signals reflect biomarker status in gliomas.Scientific Reports, 9(1):17080, 2019

Ling-Wei Kong, Jin Chen, Heng Zhao, Kun Yao, Sheng-Yu Fang, Zheng Wang, Yin-Yan Wang, and Shou-Wei Li. Intratumoral susceptibility signals reflect biomarker status in gliomas.Scientific Reports, 9(1):17080, 2019. https://doi.org/10.1038/s41598-019-53629-w

work page doi:10.1038/s41598-019-53629-w 2019
[12]

Magnetic susceptibility-based imaging in gliomas: Insights into tumor grading and margin delineation.NMR in Biomedicine, 38(10):e70140, 2025

Anita Ebrahimpour, Tayyebeh Ebrahimi, Mahboubeh Masoumbeigi, and Amin Yeganehdoust. Magnetic susceptibility-based imaging in gliomas: Insights into tumor grading and margin delineation.NMR in Biomedicine, 38(10):e70140, 2025. https://doi.org/10.1002/nbm.70140

work page doi:10.1002/nbm.70140 2025
[13]

Reichenbach

Alexander Rauscher, Jan Sedlacik, Markus Barth, Hans-Joachim Mentzel, and J¨ urgen R. Reichenbach. Magnetic susceptibility-weighted MR phase imaging of the human brain.AJNR American Journal of Neuroradiology, 26(4):736–742, 2005

2005
[14]

Murray, Carsten Rother, Ullrich K¨ othe, and Heinz-Peter Schlemmer

Jens Kleesiek, Benedikt Kersjes, Kai Ueltzh¨ offer, Jacob M. Murray, Carsten Rother, Ullrich K¨ othe, and Heinz-Peter Schlemmer. Discovering digital tumor signatures—using latent code representations to manipulate and classify liver lesions.Cancers, 13(13):3108, 2021. https://doi.org/10.3390/cancers13133108

work page doi:10.3390/cancers13133108 2021
[15]

PathologyGAN: Learning deep representations of cancer tissue.Machine Learning for Biomedical Imaging, 1(MIDL 2020 special issue):1–47, 2021

Adalberto Claudio Quiros, Roderick Murray-Smith, and Ke Yuan. PathologyGAN: Learning deep representations of cancer tissue.Machine Learning for Biomedical Imaging, 1(MIDL 2020 special issue):1–47, 2021. https://doi.org/10.59275/j.melba.2021-gfgg

work page doi:10.59275/j.melba.2021-gfgg 2020
[16]

PhaseGen: A diffusion-based approach for complex-valued MRI data generation.arXiv preprint arXiv:2504.07560, 2025

Moritz Rempe, Fabian H¨ orst, Helmut Becker, Marco Schlimbach, Lukas Rotkopf, Kevin Kr¨ oninger, and Jens Kleesiek. PhaseGen: A diffusion-based approach for complex-valued MRI data generation.arXiv preprint arXiv:2504.07560, 2025. https://doi.org/10.48550/arXiv.2504.07560

work page doi:10.48550/arxiv.2504.07560 2025
[17]

Zbontar, F

Jure Zbontar, Florian Knoll, Anuroop Sriram, Tullie Murrell, Zhengnan Huang, Matthew J. Muckley, Aaron Defazio, Ruben Stern, Patricia Johnson, Mary Bruno, Marc Parente, Krzysztof J. Geras, Joe Katsnelson, Hersh Chandarana, Zizhao Zhang, Michal Drozdzal, Adriana Romero, Michael Rabbat, Pascal Vincent, Nafissa Yakubova, James Pinkerton, Duo Wang, Erich Owen...

work page doi:10.48550/arxiv.1811.08839 2018
[18]

Lui, Michael S

Ruiyang Zhao, Burhaneddin Yaman, Yuxin Zhang, Russell Stewart, Austin Dixon, Florian Knoll, Zhengnan Huang, Yvonne W. Lui, Michael S. Hansen, and Matthew P. Lungren. fastMRI+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data.Scientific Data, 9(1):152, 2022. https://doi.org/10.1038/s41597-022-01255-z

work page doi:10.1038/s41597-022-01255-z 2022
[19]

Murphy, Patrick Virtue, Michael Elad, John M

Martin Uecker, Peng Lai, Mark J. Murphy, Patrick Virtue, Michael Elad, John M. Pauly, Shreyas S. Vasanawala, and Michael Lustig. ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA.Magnetic Resonance in Medicine, 71(3):990–1001, 2014. https://doi.org/10.1002/mrm.24751. 15

work page doi:10.1002/mrm.24751 2014
[20]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. https://doi.org/10.1109/CVPR.2016.90

work page doi:10.1109/cvpr.2016.90 2016
[21]

Learning structured output representation using deep conditional generative models

Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. InAdvances in Neural Information Processing Systems (NeurIPS), pages 3483–3491, 2015

2015
[22]

FiLM : Visual reasoning with a general conditioning layer

Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Visual reasoning with a general conditioning layer. InProceedings of the AAAI Conference on Artificial Intelligence, volume 32, pages 3942–3951, 2018. https://doi.org/10.1609/aaai.v32i1.11671

work page doi:10.1609/aaai.v32i1.11671 2018
[23]

Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InProceedings of the International Conference on Learning Representations (ICLR), 2023

2023
[24]

U-Net: convolutional networks for biomedical image segmentation , booktitle =

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. InProceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241, 2015. https://doi.org/10.1007/978-3-319-24574-4 28

work page doi:10.1007/978-3-319-24574-4 2015
[25]

Classifier-Free Diffusion Guidance

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. https://doi.org/10.48550/arXiv.2207.12598

work page internal anchor Pith review doi:10.48550/arxiv.2207.12598 2022
[26]

Generative models improve fairness of medical classifiers under distribution shifts.Nature Medicine, 30(4):1166–1173, 2024

Ira Ktena, Olivia Wiles, Isabela Albuquerque, Sylvestre-Alvise Rebuffi, Ryutaro Tanno, Abhijit Guha Roy, Shekoofeh Azizi, Danielle Belgrave, Pushmeet Kohli, Taylan Cemgil, Alan Karthikesalingam, and Sven Gowal. Generative models improve fairness of medical classifiers under distribution shifts.Nature Medicine, 30(4):1166–1173, 2024. https://doi.org/10.103...

work page doi:10.1038/s41591-024-02838-6 2024
[27]

Gamble, Hari M

Bardia Khosravi, Frank Li, Theo Dapamede, Pouria Rouzrokh, Cooper U. Gamble, Hari M. Trivedi, Cody C. Wyles, Andrew B. Sellergren, Saptarshi Purkayastha, Bradley J. Erickson, and Judy W. Gichoya. Synthetically enhanced: unveiling synthetic data’s potential in medical imaging research.eBioMedicine, 104:105174, 2024. https://doi.org/10.1016/j.ebiom.2024.105174. 16

work page doi:10.1016/j.ebiom.2024.105174 2024