Pith. sign in

REVIEW 2 major objections 5 minor 53 references

Benchmarking of Deep Learning Methods for Generic MRI Multi-Organ Abdominal Segmentation

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Across 137 MRI volumes from three public datasets, the paper finds MRSegmentator is the most accurate and generalizable open-source tool for multi-organ abdominal segmentation, and that a CT-only-trained alternative, ABDSynth, is viable…

desk verdict A solid, useful benchmark of abdominal MRI segmentation tools, with one headline claim that should be qualified: 'best' depends on how you weight Dice vs. HD95. read the letter →

arxiv 2507.17971 v2 pith:BT6HXDK3 submitted 2025-07-23 eess.IV cs.CV

Deepa Krishnaswamy , Cosmin Ciausu , Steve Pieper , Ron Kikinis , Benjamin Billot , Andrey Fedorov This is my paper · ORCID
classification eess.IVcs.CV
keywords abdominalMRIsegmentationbenchmarkMRSegmentatorABDSynthSynthSegmulti-organdomainrandomizationgeneralizability
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 aims to answer a practical question: which openly available deep-learning tool should be used to segment multiple abdominal organs in MRI scans without retraining for each new sequence or institution. It compares three existing tools, MRSegmentator, MRISegmentator-Abdomen, and TotalSegmentator MRI, and introduces a fourth, ABDSynth, trained only on CT label maps with no real MRI images. On three public datasets spanning 137 volumes, five MRI sequences, three manufacturers, and both healthy and diseased subjects, the authors report that MRSegmentator is the most accurate and most consistent, making it the recommended default for out-of-the-box use. ABDSynth trails modestly in accuracy but avoids the costly expert-annotation loop entirely, which matters when the annotation budget is scarce. The paper also releases evaluation code and the test cohort so future methods can be compared on the same footing.

What carries the argument

The load-bearing evaluation design is a three-dataset benchmark (AMOS, CHAOS, LiverHCCSeg) scored with two complementary metrics: Dice overlap and the 95th-percentile Hausdorff distance. Dice captures the fraction of the organ correctly recovered, while HD95 captures the worst typical boundary error; the two together are what make MRSegmentator's consistent spatial coherence visible and MRISegmentator-Abdomen's outliers disqualifying. For ABDSynth, the key mechanism is domain randomization: a Gaussian mixture model conditioned on CT segmentation label maps generates synthetic MRI-like volumes with randomized intensities, resolutions, contrasts, and spatial transforms, so the downstream segmentation network is forced to learn anatomy rather than any particular scanner's intensity scale. The same mechanism lets ABDSynth be trained with only 128 CT label maps and no real images.

What would settle it

Recompute the AMOS results after clipping HD95 at a clinically plausible threshold such as 10 mm; MRSegmentator's HD95 edge over MRISegmentator-Abdomen is driven by extreme outliers (for example, liver HD95 of 18.86 mm versus 2.87 mm), so if the gap largely disappears after clipping, the 'best' claim rests on outlier handling rather than typical boundary quality. A separate check is to run all four methods on a fourth institution's T1 arterial-phase liver dataset; if MRISegmentator-Abdomen wins there on both Dice and HD95, the 'most generalizable' conclusion is contradicted.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that MRSegmentator outperforms the other evaluated open-source methods in both in-domain accuracy and out-of-domain generalization for multi-organ abdominal MRI segmentation. The evidence comes from held-out cohorts the methods never saw during training: AMOS (60 diseased subjects, 15 organs), CHAOS (20 healthy subjects, three sequences), and LiverHCCSeg (17 hepatocellular-carcinoma patients with two-rater liver labels). The authors attribute MRSegmentator's advantage to the diversity of its training data, multiple T1 and T2 sequences plus CT, rather than to any architectural difference, since all three real-data methods share the same underlying architecture. They also establish that ABDSynth, though trained without real images, reaches competitive Dice on high-contrast organs such as liver, spleen, and kidneys, at lower annotation cost and with faster inference. A notable secondary finding is that MRISegmentator-Abdomen obtains the highest Dice on most AMOS organs while producing far worse HD95 values, showing that overlap and boundary quality can diverge sharply.

Load-bearing premise

The overall ranking treats Dice overlap and HD95 boundary distance as equally important; if a user cares only about overlap, MRISegmentator-Abdomen has higher Dice for the majority of AMOS organs, so calling MRSegmentator 'best' depends on that weighting.

Editorial extensions

If this is right

  • For default out-of-the-box abdominal MRI segmentation, MRSegmentator is the model to start with whenever both overlap and boundary coherence matter.
  • ABDSynth shows that a practical multi-organ MRI segmenter can be built from CT-only annotations, so annotation budgets can be redirected when MRI labels are scarce.
  • The dissociation between Dice and HD95 observed for MRISegmentator-Abdomen means single-metric evaluation is insufficient; future benchmarks should report both or use a boundary-aware composite.
  • On CHAOS, models trained on the evaluated sequences generalized best, and MRISegmentator-Abdomen's drop there shows that sequence coverage and resolution diversity in training are decisive for cross-sequence generalization.
  • The released evaluation code and test datasets give future methods a fixed, externally comparable baseline rather than a moving target.

Reading between the lines

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

  • If this pattern holds beyond the tested cohorts, domain-randomization training may make MRI segmentation feasible in low-resource settings where CT annotations exist but expert MRI annotation does not, at the price of accepting lower accuracy on small and deformable organs.
  • The inter-rater HD95 on LiverHCCSeg (15.7 mm) is worse than every automated method, which suggests that near the top of the ranking the differences between methods may be comparable to annotation noise; a pooled multi-rater reference standard could change the ordering.
  • The results imply that generalizability is not one property but several, robustness to sequence, resolution, and pathology, and that a method can lead on one axis while losing on another; a method's rank may therefore shift with the composition of the test set.
  • A natural extension would be to train ABDSynth with a larger and more diverse set of CT segmentations or with augmentation targeted at T1 arterial-phase contrast, and to test whether its gap to MRSegmentator on organs like the pancreas and duodenum closes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The manuscript benchmarks four open-source deep learning methods for multi-organ abdominal MRI segmentation: MRSegmentator, MRISegmentator-Abdomen, TotalSegmentator MRI, and ABDSynth (a new SynthSeg-based model trained purely on CT segmentations). The evaluation uses three public datasets not seen during training (AMOS, CHAOS, LiverHCCSeg), covering multiple manufacturers, sequences, and subject conditions, with Dice and HD95 as metrics. The authors report that MRSegmentator achieves the best overall performance and generalizability, while ABDSynth is slightly less accurate but requires no manual MRI annotations. The paper also releases evaluation code and the ABDSynth model weights.

Significance. If the benchmark's conclusions are accepted, the paper provides a valuable independent comparison of currently available open-source tools for abdominal MRI segmentation, a clinically important but under-benchmarked area. The public release of evaluation code and the ABDSynth weights is a concrete contribution to reproducibility. The demonstration that a model trained only on CT segmentations can achieve competitive performance on MRI is an interesting finding with practical implications for annotation-scarce settings. However, the headline claim that MRSegmentator is the 'best' method depends on how the two metrics, Dice and HD95, are weighted; the manuscript does not currently operationalize this choice, which weakens the central conclusion.

major comments (2)
  1. [Section III-A, Table III] The statement that MRSegmentator 'achieves the best performance' is not directly supported by the reported results because no composite metric or pre-specified weighting of Dice and HD95 is defined. On AMOS, MRISegmentator-Abdomen achieves higher Dice for the majority of organs (average gap of 0.035 over MRSegmentator, as stated in Section III-A), while MRSegmentator exhibits substantially better HD95 and fewer outliers. The paper does not justify an implicit equal weighting of the two metrics. Since user priorities may vary, the conclusion should either be based on a clearly defined composite metric (e.g., average of ranks or normalized scores) or be qualified to state that MRSegmentator is recommended when boundary accuracy (HD95) is prioritized, while MRISegmentator-Abdomen may be preferable when overlap (Dice) is the primary concern on T1-weighted sequences. This issue is load-bearing because the abstract's unqualified 'best performance' claim rests on it.
  2. [Section III, Table III caption] The statistical testing procedure is under-specified. The caption says asterisks denote statistical significance with all other models at the 5% level using a Bonferroni-corrected Wilcoxon signed-rank test, but the manuscript does not describe how the pairwise comparisons were set up: whether the correction is applied across the number of methods, regions, or both; whether tests are one-sided or two-sided; and how missing values or ties are handled. The asterisks are used to substantiate the ranking claims, especially in the abstract and discussion, so a complete description of the multiple-comparison procedure is needed. Without this, the reader cannot assess the strength of the evidence for MRSegmentator's superiority.
minor comments (5)
  1. [Appendix, Algorithm 1] Algorithm 1 is ambiguous: as written, it appears to loop over all possible cluster counts and fit a GMM for each (for KBG in Rand{3,4,5,6,7} and for KFG in Rand{1,2,3}), whereas the text describes randomly sampling a single cluster count at runtime. The pseudocode should be revised to match the described procedure, for example by sampling a value first and then fitting one GMM.
  2. [Table I] The resolution ranges for some training datasets are reported as very wide ranges (e.g., TotalSegmentator MRI: [0.29×0.29×0.43] to [7.50×25.00×28.0]); it would be helpful to clarify whether these are full ranges or mean values, and to add a footnote explaining the notation for voxel dimensions.
  3. [Section II-B, Table II] For the AMOS dataset, the sequences are listed as 'MRI (sequences not provided)'; specifying the available sequence types, even approximately, would be useful for interpreting the generalizability results and comparing with the other datasets.
  4. [Figures 4 and 5] The notation for CHAOS sequences varies between 'T1 in-phase/out-phase' in the text and 'T1 D-IP/T1 D-OP' in figures; please standardize the terminology throughout the manuscript.
  5. [Section V] The conclusion notes that TotalVibeSegmentator was not evaluated; it would also be useful to explicitly mention that the compared methods have different label sets and that only the intersection of regions is benchmarked, as this limits the scope of the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: external benchmark with fixed pretrained models; no fitted parameters hidden in the evaluation.

full rationale

The paper's central claim is an empirical ranking of four fixed segmentation models on three external public datasets (AMOS, CHAOS, LiverHCCSeg) that are not used in training any evaluated method. The conclusion that MRSegmentator performs best is derived from reported Dice/HD95 statistics in Table III and associated statistical tests, not from a definition or from parameters fitted to those datasets. ABDSynth is introduced by the authors, but it is evaluated on the same held-out external datasets and is not claimed to outperform the other methods; its training (128 CT segmentations, SynthSeg GMM generation) and the SynthSeg citation [32] are construction details, not a load-bearing uniqueness or equivalence argument. The comparison is therefore self-contained against external benchmarks, and any concern about metric weighting (Dice vs HD95) is a correctness or interpretability issue, not circularity. The self-citation of SynthSeg by co-author Billot is not load-bearing because the benchmark conclusion does not depend on the cited method's validity; SynthSeg is code-reproduced and its assumptions do not include the target ranking. Score 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central benchmark claim does not rest on fitted parameters. The listed values are design choices for ABDSynth, inherited or slightly expanded from SynthSeg [32]; they were not tuned on the evaluation datasets, so they do not constitute hidden fitting. The important domain assumptions are the representativeness of the three public test sets and the transfer of SynthSeg-style domain randomization from brain and cardiac to abdominal MRI.

free parameters (3)
  • ABDSynth label subclustering counts = foreground {1,2,3}, background {3,4,5,6,7}
    Chosen by hand in the Appendix to improve synthetic contrast realism; not fitted to the evaluation data, but affects ABDSynth performance.
  • ABDSynth arm-removal probability = 0.5
    Chosen by hand to simulate trunk-only MRI acquisitions; affects synthetic pose diversity.
  • ABDSynth training iterations = 500,000
    Chosen to match SynthSeg's training schedule; not fitted to the benchmark.
assumptions (3)
  • domain assumption The three public evaluation datasets (AMOS, CHAOS, LiverHCCSeg) are representative of the generic MRI abdominal segmentation landscape, and their ground truth annotations are of sufficient quality to rank methods.
    Introduced in Section II-B; the validity of the ranking depends on these datasets being a fair and adequate testbed.
  • domain assumption SynthSeg's domain randomization can generate synthetic MRI-like images that are sufficiently realistic to train a network that generalizes to real MRI.
    Invoked in Section II-A4 as the core premise behind ABDSynth; if this transfer fails, ABDSynth's performance would not be meaningful.
  • standard math Dice and HD95 are appropriate metrics for segmentation quality.
    Stated in Section II-C; the ranking of methods is directly determined by these two metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking of Deep Learning Methods for Generic MRI Multi-Organ Abdominal Segmentation." pith.science (2026). https://pith.science/paper/BT6HXDK3

@misc{pith2026250717971,
  author       = {Pith},
  title        = {Pith review of: Benchmarking of Deep Learning Methods for Generic MRI Multi-Organ Abdominal Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BT6HXDK3}},
  note         = {Machine review of arXiv:2507.17971}
}
read the original abstract

Recent advances in deep learning have led to robust automated tools for segmentation of abdominal computed tomography (CT). Meanwhile, segmentation of magnetic resonance imaging (MRI) is substantially more challenging due to the inherent signal variability and the increased effort required for annotating training datasets. Hence, existing approaches are trained on limited sets of MRI sequences, which might limit their generalizability. To characterize the landscape of MRI abdominal segmentation tools, we present here a comprehensive benchmarking of the three state-of-the-art and open-source models: MRSegmentator, MRISegmentator-Abdomen, and TotalSegmentator MRI. Since these models are trained using labor-intensive manual annotation cycles, we also introduce and evaluate ABDSynth, a SynthSeg-based model purely trained on widely available CT segmentations (no real images). More generally, we assess accuracy and generalizability by leveraging three public datasets (not seen by any of the evaluated methods during their training), which span all major manufacturers, five MRI sequences, as well as a variety of subject conditions, voxel resolutions, and fields-of-view. Our results reveal that MRSegmentator achieves the best performance and is most generalizable. In contrast, ABDSynth yields slightly less accurate results, but its relaxed requirements in training data make it an alternative when the annotation budget is limited. The evaluation code and datasets are given for future benchmarking at https://github.com/deepakri201/AbdoBench, along with inference code and weights for ABDSynth.

Figures

Figures reproduced from arXiv: 2507.17971 by the authors.

Figure 1
Figure 1. Overview of an ABDSynth training step. 1) A CT segmentation is sampled from the training set. 2) A synthetic volume is generated using a segmentation-conditioned GMM with randomized parameters. 3,4) Abdominal volume/segmentation pairs are used to train a supervised 3D U-Net. TABLE II SUMMARY OF THE PUBLICLY AVAILABLE DATASETS USED FOR EVALUATION. BRACKETS DENOTE RANGES. Dataset # subj. Manufacturer Sequence Presence… view at source ↗
Figure 2
Figure 2. Dice (top) and HD95 (bottom) boxplots for AMOS results for the four benchmarked methods. We observe similar performances for all methods across the liver, spleen, and kidneys, but highly variable results across the regions that are smaller and/or with more variable morphologies. (a) (b) Ground truth MRSegmentator MRISegmentator-Abdomen TotalSegmentator MRI ABDSynth [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Sample segmentations by all methods on AMOS. (a) Pancreas slice, where all methods do not fully segment the region. (b) 3D renderings for another subject. Arrows indicate segmentation errors in the region of corresponding color. among all methods and for almost all sequence types and evaluated regions. TotalSegmentator MRI and ABDSynth also yield fairly high Dice scores and low HD95 values. In contrast, MRISegmentat… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Dice score (left), HD95 (middle), and volume repeatability (right) obtained on CHAOS for two representative regions (liver and right kidney) across different sequences. In the volume repeatability subfigure, the consecutive points represent T1 dual in-phase, T1 dual ou…
Figure 5
Figure 5. Figure 5: CHAOS subject 39, where blue = liver and yellow = right kidney. Blue and yellow arrows point at major differences between ground truth and automated segmentations for the liver and right kidney, respectively. kidney, where all methods obtain tighter HD95 distributions …
Figure 6
Figure 6. Figure 6: Example of liver segmentations for the superior, middle, and inferior axial slices of a representative LiverHCCSeg subject. For the inferior axial slice, all methods do not segment the liver well, likely due to the presence of a hepatocellular carcinoma, as indicated b…
Figure 7
Figure 7. Figure 7: Right kidney segmentations for CHAOS subject 1. The yellow arrow points to major differences between ground truth and automated segmentations, where the ground truth includes the renal pelvis. note that the public datasets used here for benchmarking do not provide subj…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages

  1. [1]

    MRSegm en- tator: Robust multi-modality segmentation of 40 classes in MRI and CT sequences,

    H. H¨ antze, L. Xu, F. Dorfner, L. Donle, D. Truhn, H. Aerts, M. Prokop, B. van Ginneken, A. Hering, L. Adams, and K. Bressem, “MRSegm en- tator: Robust multi-modality segmentation of 40 classes in MRI and CT sequences,” arXiv:2405.06463, 2024. 7https://github.com/PerkLab/SlicerSandbox#remove-ct- table Algorithm 1 Background and foreground labels clusteri...

    arXiv 2024

  2. [2]

    UK bi obank: an open access resource for identifying the causes of a wide r ange of complex diseases of middle and old age,

    C. Sudlow, J. Gallacher, N. Allen, V . Beral, P . Burton, J. Dan esh, P . Downey, P . Elliott, J. Green, M. Landray, and B. Liu, “UK bi obank: an open access resource for identifying the causes of a wide r ange of complex diseases of middle and old age,” PLoS medicine, vol. 12, no. 3, p. e1001779, 2015

    work page 2015

  3. [3]

    TotalSegme ntator: Robust segmentation of 104 anatomic structures in CT images ,

    J. Wasserthal, H. Breit, M. Meyer, M. Pradella, D. Hinck, A. S auter, T. Heye, D. Boll, J. Cyriac, S. Y ang, and M. Bach, “TotalSegme ntator: Robust segmentation of 104 anatomic structures in CT images ,” Radi- ology: Artificial Intelligence , vol. 5, no. 5, p. 102789, 2023

    work page 2023

  4. [4]

    Dataset with segmentations of 104 importan t anatomical structures in 1204 CT images,

    J. Wasserthal, “Dataset with segmentations of 104 importan t anatomical structures in 1204 CT images,” 2023. [Online]. A vailable: https://zenodo.org/records/6802614

    work page arXiv 2023

  5. [5]

    MRISegmentator-Abdomen: A full y automated multi-organ and structure segmentation tool for T1-weighted abdominal MRI,

    Y . Zhuang, T. Mathai, P . Mukherjee, B. Khoury, B. Kim, B. Hou, N. Rabbee, and R. Summers, “MRISegmentator-Abdomen: A full y automated multi-organ and structure segmentation tool for T1-weighted abdominal MRI,” arXiv.2405.05944, 2024

    arXiv 2024

  6. [6]

    TotalSegmentator MRI: Sequence-independent segmen tation of 59 anatomical structures in MR images,

    T. D’Antonoli, L. Berger, A. Indrakanti, N. Vishwanathan, J . Weiß, M. Jung, Z. Berkarda, A. Rau, M. Reisert, T. K¨ ustner, and A. W al- ter, “TotalSegmentator MRI: Sequence-independent segmen tation of 59 anatomical structures in MR images,” arXiv:2405.19492, 2024

    arXiv 2024

  7. [7]

    N a- tional cancer institute imaging data commons: Toward trans parency, reproducibility, and scalability in imaging artificial int elligence,

    A. Fedorov, W. Longabaugh, D. Pot, D. Clunie, S. Pieper, D. Gi bbs, C. Bridge, M. Herrmann, A. Homeyer, R. Lewis, and H. Aerts, “N a- tional cancer institute imaging data commons: Toward trans parency, reproducibility, and scalability in imaging artificial int elligence,” Ra- diographics, vol. 43, no. 12, p. e230180, 2023

    work page 2023

  8. [8]

    Auto mated segmentation of tissues using CT and MRI: a systematic revie w,

    L. Lenchik, L. Heacock, A. Weaver, R. Boutin, T. Cook, J. Itri, C. Filippi, R. Gullapalli, J. Lee, M. Zagurovskaya, and T. Retson, “Auto mated segmentation of tissues using CT and MRI: a systematic revie w,” Academic radiology, vol. 26, no. 12, pp. 1695–706, 2019

    work page 2019

Show all 53 references
  1. [9]

    Pancreas volume in health and disease: a systematic review and meta- analysis,

    S. DeSouza, R. Singh, H. Y oon, R. Murphy, L. Plank, and M. Petr ov, “Pancreas volume in health and disease: a systematic review and meta- analysis,” Expert review of gastroenterology & hepatology, vol. 12, no. 8, pp. 757–66, 2018

    2018

  2. [10]

    Assessment of kidney volumes from mri: acquisit ion and segmentation techniques,

    F. Z ¨ ollner, E. Svarstad, A. Munthe-Kaas, L. Schad, A. Lunde rvold, and J. Rørvik, “Assessment of kidney volumes from mri: acquisit ion and segmentation techniques,” American Journal of Roentgenology, vol. 199, no. 5, pp. 1060–9, 2012

    2012

  3. [11]

    The value of m agnetic resonance imaging for radiotherapy planning,

    P . Dirix, K. Haustermans, and V . V andecaveye, “The value of m agnetic resonance imaging for radiotherapy planning,” In Seminars in radiation oncology, vol. 24, no. 3, pp. 151–159, 2014

    2014

  4. [12]

    Integrated MRI-guided radiotherapy—opportunities and c hallenges,

    P . Keall, C. Brighi, C. Glide-Hurst, G. Liney, P . Liu, S. Lydi ard, C. Paganelli, T. Pham, S. Shan, A. Tree, and U. van der Heide, “Integrated MRI-guided radiotherapy—opportunities and c hallenges,” Nature Reviews Clinical Oncology , vol. 19, no. 7, pp. 458–70, 2022

    2022

  5. [13]

    Radiomics: images ar e more than pictures, they are data,

    R. Gillies, P . Kinahan, and H. Hricak, “Radiomics: images ar e more than pictures, they are data,” Radiology, vol. 278, no. 2, pp. 563–77, 2016

    2016

  6. [14]

    CT and MRI of abdominal can cers: current trends and perspectives in the era of radiomics and a rtificial 10 IEEE TRANSACTIONS ON MEDICAL IMAGING intelligence,

    M. Barat, A. Pellat, C. Hoeffel, A. Dohan, R. Coriat, E. Fishm an, S. Nougaret, L. Chu, and P . Soyer, “CT and MRI of abdominal can cers: current trends and perspectives in the era of radiomics and a rtificial 10 IEEE TRANSACTIONS ON MEDICAL IMAGING intelligence,” Japanese jour...

    2024

  7. [15]

    Recommendations for MRI-based contouring of gross tumor v olume and organs at risk for radiation therapy of pancreatic cance r,

    H. Heerkens, W. Hall, X. Li, P . Knechtges, E. Dalah, E. Paulso n, C. van den Berg, G. Meijer, E. Koay, C. Crane, and K. Aitken, “Recommendations for MRI-based contouring of gross tumor v olume and organs at risk for radiation therapy of pancreatic cance r,” Practical radiatio...

    2017

  8. [16]

    vOARiability: Interobserver and intermodality variabil ity analysis in oar contouring from head and neck CT and MR images,

    G. Podobnik, B. Ibragimov, P . Peterlin, P . Strojan, and T. Vr tovec, “vOARiability: Interobserver and intermodality variabil ity analysis in oar contouring from head and neck CT and MR images,” Medical Physics, vol. 51, no. 3, pp. 2175–86, 2024

    2024

  9. [17]

    U-Net: Convolutio nal net- works for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-Net: Convolutio nal net- works for biomedical image segmentation,” In Medical Image Com- puting and Computer-Assisted Intervention–MICCAI 2015: 1 8th Inter- national Conference, Munich, Germany, Proceedings, Part I II 18 , pp. 234–241, 2015

    2015

  10. [18]

    A learning strategy for contrast-agnostic mri se gmentation,

    B. Billot, K. Greve, D.and V an Leemput, B. Fischl, J. Iglesia s, and A. Dalca, “A learning strategy for contrast-agnostic mri se gmentation,” Medical Imaging with Deep Learning (MIDL) , p. 175–93, 2020

    2020

  11. [19]

    A survey on transfer learning,

    S. Pan and Q. Y ang, “A survey on transfer learning,” IEEE Transactions on Knowledge and Data Engineering , vol. 22, p. 45–59, 2010

    2010

  12. [20]

    Fully automated multiorgan segmentation in abdom- inal magnetic resonance imaging with deep neural networks,

    Y . Chen, D. Ruan, J. Xiao, L. Wang, B. Sun, R. Saouaf, W. Y ang, D. Li, and Z. Fan, “Fully automated multiorgan segmentation in abdom- inal magnetic resonance imaging with deep neural networks, ” Medical physics, vol. 47, no. 10, pp. 4971–82, 2020

    2020

  13. [21]

    Deep learning-bas ed auto- mated abdominal organ segmentation in the uk biobank and ger man national cohort magnetic resonance imaging studies,

    T. Kart, M. Fischer, T. K¨ ustner, T. Hepp, F. Bamberg, S. Winz eck, B. Glocker, D. Rueckert, and S. Gatidis, “Deep learning-bas ed auto- mated abdominal organ segmentation in the uk biobank and ger man national cohort magnetic resonance imaging studies,” Investigative Ra- dio...

    2021

  14. [22]

    Abdomennet: deep neural network for abdomin al organ segmentation in epidemiologic imaging studies,

    A. Rickmann, J. Senapati, O. Kovalenko, A. Peters, F. Bamber g, and C. Wachinger, “Abdomennet: deep neural network for abdomin al organ segmentation in epidemiologic imaging studies,” BMC medical imaging, vol. 22, no. 1, p. 168, 2022

    2022

  15. [23]

    Deep learning auto-segmentation on multi- sequence magnetic resonance images for upper abdominal org ans,

    A. Amjad, J. Xu, D. Thill, Y . Zhang, J. Ding, E. Paulson, W. Hal l, B. Erickson, and X. Li, “Deep learning auto-segmentation on multi- sequence magnetic resonance images for upper abdominal org ans,” Frontiers in oncology , vol. 13, pp. 1–12, 2023

    2023

  16. [24]

    nnU-Net: a self-configuring method for deep learning-based biomedic al image segmentation,

    F. Isensee, P . Jaeger, S. Kohl, J. Petersen, and K. Maier-Hei n, “nnU-Net: a self-configuring method for deep learning-based biomedic al image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–11, 2021

    2021

  17. [25]

    The german national cohort: aims, st udy design and organization,

    G. N. C. G. Consortium, “The german national cohort: aims, st udy design and organization,” European journal of epidemiology , vol. 29, no. 5, pp. 371–82, 2014

    2014

  18. [26]

    Lymph node detection in t2 mri with transformers,

    T. Mathai, S. Lee, D. Elton, T. Shen, Y . Peng, Z. Lu, and R. Summ ers, “Lymph node detection in t2 mri with transformers,” In Medical imaging 2022: computer-aided diagnosis, SPIE. , vol. 12033, pp. 869–873, 2022

    2022

  19. [27]

    Universal l ymph node detection in t2 mri using neural networks,

    T. Mathai, S. Lee, T. Shen, Z. Lu, and R. Summers, “Universal l ymph node detection in t2 mri using neural networks,” International journal of computer assisted radiology and surgery , vol. 18, no. 2, pp. 313–8, 2023

    2023

  20. [28]

    Domain shift in computer vision models for M RI data analysis: an overview,

    E. Kondrateva, M. Pominova, E. Popova, M. Sharaev, A. Bernst ein, and E. Burnaev, “Domain shift in computer vision models for M RI data analysis: an overview,” in Thirteenth International Conference on Machine Vision, vol. 11605. SPIE, 2021, pp. 126–133

    2021

  21. [29]

    A 3d unsupervised domain adaptation framework combining style translation a nd self- training for abdominal organs segmentation,

    J. Li, Q. Chen, H. Ding, H. Liu, and L. Wan, “A 3d unsupervised domain adaptation framework combining style translation a nd self- training for abdominal organs segmentation,” in Fast and Low-Resource Semi-supervised Abdominal Organ Segmentation . MICCAI, 2024, pp. 209–224

    2024

  22. [30]

    Unsupervi sed domain adaptation for abdominal organ segmentation using p seudo labels and organ attention cyclegan,

    J. Wu, G. Zhang, X. Qi, H. Wang, X. Liu, and G. Wang, “Unsupervi sed domain adaptation for abdominal organ segmentation using p seudo labels and organ attention cyclegan,” in Fast and Low-Resource Semi- supervised Abdominal Organ Segmentation . MICCAI, 2024, pp. 225– 242

    2024

  23. [31]

    Unpaired image -to- image translation using cycle-consistent adversarial net works,

    J.-Y . Zhu, T. Park, P . Isola, and A. A. Efros, “Unpaired image -to- image translation using cycle-consistent adversarial net works,” in IEEE International Conference on Computer Vision (ICCV) , 2017

    2017

  24. [32]

    SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining,

    B. Billot, D. Greve, O. Puonti, A. Thielscher, K. V an Leemput, B. Fischl, A. Dalca, and J. Iglesias, “SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining,” Medical image analysis, vol. 86, p. 102789, 2023

    2023

  25. [33]

    Robust machine learning segmentation for large-scale ana lysis of heterogeneous clinical brain mri datasets,

    B. Billot, C. Magdamo, Y . Cheng, S. E. Arnold, S. Das, and J. E. Iglesias, “Robust machine learning segmentation for large-scale ana lysis of heterogeneous clinical brain mri datasets,” Proceedings of the National Academy of Sciences , vol. 120, no. 9, 2023

    2023

  26. [34]

    Domain randomization for transferring deep neural networ ks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P . Abb eel, “Domain randomization for transferring deep neural networ ks from simulation to the real world,” IEEE/RSJ international conference on intelligent robots and systems (IROS) , pp. 23–30, 2017

    2017

  27. [35]

    AMOS: A large-scale abdominal mult i- organ benchmark for versatile medical image segmentation,

    Y . Ji, H. Bai, C. Ge, J. Y ang, Y . Zhu, R. Zhang, Z. Li, L. Zhang, W. Ma, X. Wan, and P . Luo, “AMOS: A large-scale abdominal mult i- organ benchmark for versatile medical image segmentation, ” Advances in Neural Information Processing Systems , vol. 35, pp. 36 722–32, 2022

    2022

  28. [36]

    CHAOS - combined (CT-MR) healthy abdominal organ segmentation cha llenge data,

    A. Kavur, A. Selver, O. Dicle, M. Barıs ¸, and N. Gezer, “CHAOS - combined (CT-MR) healthy abdominal organ segmentation cha llenge data,” 2019. [Online]. Available: https://zenodo.org/records/3431873

    2019

  29. [37]

    Comparison of semi- automatic and deep learning-based automatic methods for li ver segmen- tation in living liver transplant donors,

    A. Kavur, N. Gezer, M. Barıs ¸, Y . S ¸ ahin, S.¨Ozkan, B. Baydar, U. Y¨ uksel, c. Kılıkc ¸ıer, c. Olut, G. Akar, and G. ¨Unal, “Comparison of semi- automatic and deep learning-based automatic methods for li ver segmen- tation in living liver transplant donors,” Diagnostic and...

    2020

  30. [38]

    CHAOS challenge- combined (CT-MR) healthy abdominal organ segmentation,

    A. Kavur, N. Gezer, M. Barıs ¸, S. Aslan, P . Conze, V . Groza, D. Pham, S. Chatterjee, P . Ernst, S. ¨Ozkan, and B. Baydar, “CHAOS challenge- combined (CT-MR) healthy abdominal organ segmentation,” Medical Image Analysis , vol. 69, p. 101950, 2021

    2021

  31. [39]

    Liver- HCCSeg: A publicly available multiphasic MRI dataset with l iver and HCC tumor segmentations and inter-rater agreement analysi s,

    M. Gross, S. Arora, S. Huber, A. K¨ uc¨ ukkaya, and J. Onofrey, “Liver- HCCSeg: A publicly available multiphasic MRI dataset with l iver and HCC tumor segmentations and inter-rater agreement analysi s,” Data in Brief, vol. 51, p. 109662, 2023

    2023

  32. [40]

    L iverHC- Ceg: A publicly available multiphasic MRI dataset with live r and HCCC tumor segmentations and inter-rater agreement analysis,

    M. Gross, S. Arora, S. Huber, A. Kucukkaya, and J. Onofrey, “L iverHC- Ceg: A publicly available multiphasic MRI dataset with live r and HCCC tumor segmentations and inter-rater agreement analysis,” Zenodo, 2023

    2023

  33. [41]

    Monai label: A fra mework for ai-assisted interactive labeling of 3D medical images,

    A. Diaz-Pinto, S. Alle, V . Nath, Y . Tang, A. Ihsani, M. Asad, F . P´ erez- Garc´ ıa, P . Mehta, W. Li, M. Flores, H. Roth, T. V ercauteren, D. Xu, P . Dogra, S. Ourselian, A. Feng, and C. MJ, “Monai label: A fra mework for ai-assisted interactive labeling of 3D medical image...

    2024

  34. [42]

    Segment ation of pelvic structures in T2 MRI via MR-to-CT synthesis,

    Y . Zhuang, T. Mathai, P . Mukherjee, and R. Summers, “Segment ation of pelvic structures in T2 MRI via MR-to-CT synthesis,” Computerized Medical Imaging and Graphics , vol. 112, p. 102335, 2024

    2024

  35. [43]

    Fully convolutional neural netw orks for volumetric medical image segmentation,

    A. S. Milletari F, Navab N, “Fully convolutional neural netw orks for volumetric medical image segmentation,” In 2016 fourth international conference on 3D vision (3DV) , pp. 65–571, 2016

    2016

  36. [44]

    Jetstream2: Accelerating cloud computing via j etstream

    D. Hancock, J. Fischer, J. Lowe, W. Snapp-Childs, M. Pierce, S. Marru, J. Coulter, M. V aughn, B. Beck, N. Merchant, E. Skidmore, and G. Jacobs, “Jetstream2: Accelerating cloud computing via j etstream.” pp. 1–8, 2021

    2021

  37. [45]

    ACC ESS: Advancing innovation: NSF’s advanced cyberinfrastructur e coordination ecosystem: Services & support,

    T. Boerner, S. Deems, T. Furlani, S. Knuth, and J. Towns, “ACC ESS: Advancing innovation: NSF’s advanced cyberinfrastructur e coordination ecosystem: Services & support,” 2023

    2023

  38. [46]

    The cancer genome atlas liver hepatocellular carcinoma collec tion (TCGA-LIHC) (version 5) [data set],

    B. Erickson, S. Kirk, Y . Lee, O. Bathe, M. Kearns, C. Gerdes, K. Rieger-Christ, and J. Lemmerman, “The cancer genome atlas liver hepatocellular carcinoma collec tion (TCGA-LIHC) (version 5) [data set],” 2016. [Online]. Avail able: https://wiki.cancerimagingarchive.net/pages/v...

    2016

  39. [47]

    Measures of the amount of ecologic association bet ween species,

    L. Dice, “Measures of the amount of ecologic association bet ween species,” Ecology, vol. 26, no. 3, pp. 297–302, 1945

    1945

  40. [48]

    Compari ng images using the hausdorff distance,

    D. Huttenlocher, G. Klanderman, and W. Rucklidge, “Compari ng images using the hausdorff distance,” IEEE Transactions on pattern analysis and machine intelligence, vol. 15, no. 9, pp. 850–63, 1993

    1993

  41. [49]

    Segmentatio n precision of abdominal anatomy for mri-based radiotherapy ,

    C. Noel, F. Zhu, A. Lee, H. Y anle, and P . Parikh, “Segmentatio n precision of abdominal anatomy for mri-based radiotherapy ,” Medical Dosimetry, vol. 39, no. 3, pp. 212–7, 2014

    2014

  42. [50]

    Pancreas segmenta tion in mri using graph-based decision fusion on convolutional n eural net- works,

    J. Cai, L. Lu, Y . Xie, F. Xing, and L. Y ang, “Pancreas segmenta tion in mri using graph-based decision fusion on convolutional n eural net- works,” In Medical Image Computing and Computer Assisted Interven- tion - MICCAI 2017: 20th International Conference , pp. 674–682, 2017

    2017

  43. [51]

    Noise and the reality gap: The use of simulation in evolutionary robotics,

    N. Jakobi, P . Husbands, and I. Harvey, “Noise and the reality gap: The use of simulation in evolutionary robotics,” in European conference on artificial life , 1995, pp. 704–720

    1995

  44. [52]

    To- talvibesegmentator: Full body mri segmentation for the nak o and uk biobank,

    R. Graf, P . Platzek, E. Riedel, C. Ramsch¨ utz, S. Starck, H. M ¨ oller, M. Atad, H. V¨ olzke, R. B¨ ulow, C. Schmidt, and J. R¨ udebusch , “To- talvibesegmentator: Full body mri segmentation for the nak o and uk biobank,” arXiv:2406.00125, 2024

    2024 arXiv

  45. [53]

    Maximum likelihood fro m incomplete data via the EM algorithm,

    A. Dempster, N. Laird, and D. Rubin, “Maximum likelihood fro m incomplete data via the EM algorithm,” Journal of the royal statistical society: series B (methodological) , vol. 39, no. 1, pp. 1–22, 1977

    1977

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.