pith. sign in

arxiv: 2605.16572 · v1 · pith:IRQ4HQBMnew · submitted 2026-05-15 · 💻 cs.CV

TriALS: Triphasic-Aided Liver Lesion Segmentation Benchmark in Non-Contrast CT

Marawan Elbatel , Mohamed Ghonim , Jiaji Mao , Zhuosheng Lin , Katharina Eckstein , Andr\'es Mart\'inez Mora , Jonathan Deissler , Maximilian Rokuss
show 57 more authors
Constantin Ulrich Zdravko Marinov Wenhui Deng Baoxun Li Huijun Hu Jun Shen Mohanad Ghonim Khadiga Omar Nassar Mariam Elbakry Menna Dyab Amr Muhammad Abdo Salem Nouran Elghitany Noha Elghitany Yi Qin Xuanqi Huang Haonan Wang Shao-Woo Yen Ahmed Elghamry Saba Salma Ahmad Xinyan Fang Jiahao Zhang Xiaodi Wang Xinghua Ma Gongning Luo Jessica C. Delmoral Jo\~ao Manuel R.S. Tavares Ankan Deria Adinath Dukre Yutong Xie Imran Razzak Dongwook Kim Matthew Choi Hanxiao Zhang Minghui Zhang Xin You Abdul Qayyum Steven A. Niederer Moona Mazher Rachika E. Hamadache Ricardo Montoya-del-Angel Robert Mart\'i Xavier Llad\'o Toufiq Musah Livingstone Eli Ayivor Enrique Almar-Munoz Agnes Mayr Kaouther Mouheb Esther E. Bron Stefan Klein Ahmed Abouelhoda Amira Adel Susan Adil Ali Rainer Stiefelhagen Klaus H. Maier-Hein Fabian Isensee Aya Yassin Xiaomeng Li
This is my paper

Pith reviewed 2026-05-20 18:51 UTC · model grok-4.3

classification 💻 cs.CV
keywords liver lesion segmentationnon-contrast CTtriphasic CTmedical image segmentationbenchmark datasetexternal validationdeep learning
0
0 comments X

The pith

A new benchmark shows liver lesion segmentation on non-contrast CT reaches only 0.57 Dice even with triphasic training data.

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

The paper creates the TriALS benchmark to measure progress on automated liver lesion segmentation when contrast agents are unavailable, a common constraint in low-resource settings. It supplies a multi-centre collection of 150 cases with full four-phase CT acquisitions and runs an evaluation on 70 held-out cases including an external cohort. Top algorithms hit 0.754 Dice on venous-phase images yet fall to 0.57 on non-contrast scans, while still beating generic models by up to 28 percent Dice on external non-contrast data. Performance tracks most closely with training-set size and pre-training choices, but a cross-year comparison indicates that simply scaling pre-training does not close the remaining gap on non-contrast images.

Core claim

The TriALS challenge establishes a standardized, multi-centre benchmark for liver lesion segmentation under contrast-limited conditions. Using 600 volumes from 150 cases acquired at Egyptian and Chinese sites, the evaluation shows that leading methods achieve near human-level accuracy on venous-phase images but suffer a substantial drop on non-contrast CT, with external validation confirming gains of up to 28 percent Dice over off-the-shelf models. Algorithm success is driven primarily by data scale and pre-training strategy, yet a persistent perceptual barrier on non-contrast scans remains even after these factors are optimized.

What carries the argument

The TriALS benchmark dataset, consisting of 150 multi-centre cases with four-phase CT acquisitions that enable triphasic-aided training for non-contrast segmentation evaluation.

If this is right

  • Training data volume and choice of pre-training directly determine how well a method performs on non-contrast liver lesion segmentation.
  • Methods tuned on the TriALS data generalize better to unseen institutions than standard pre-trained models on non-contrast scans.
  • A performance ceiling on non-contrast images persists that cannot be removed by increasing pre-training scale alone.
  • The benchmark supplies a common reference that lets groups in different regions compare progress under realistic contrast constraints.

Where Pith is reading between the lines

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

  • Future architectures may need explicit mechanisms for handling the reduced lesion conspicuity that occurs without contrast rather than relying solely on more data.
  • The same triphasic-to-non-contrast transfer pattern could be tested on other abdominal organs to see whether the observed perceptual barrier is liver-specific.
  • If the gap remains after larger-scale pre-training, synthetic contrast generation or unpaired image translation techniques become higher-priority research directions.
  • The external-validation gain suggests that releasing the full dataset and code could accelerate method development in settings without access to contrast agents.

Load-bearing premise

The 150-case multi-centre collection from Egyptian and Chinese hospitals adequately captures the range of clinical variability found in low-resource settings across Africa and Asia.

What would settle it

A new method trained only on non-contrast data that exceeds 0.65 mean Dice on the external NCCT test set without any triphasic volumes.

Figures

Figures reproduced from arXiv: 2605.16572 by Abdul Qayyum, Adinath Dukre, Agnes Mayr, Ahmed Abouelhoda, Ahmed Elghamry Saba, Amira Adel, Amr Muhammad Abdo Salem, Andr\'es Mart\'inez Mora, Ankan Deria, Aya Yassin, Baoxun Li, Constantin Ulrich, DongWook Kim, Enrique Almar-Munoz, Esther E. Bron, Fabian Isensee, Gongning Luo, Hanxiao Zhang, Haonan Wang, Huijun Hu, Imran Razzak, Jessica C. Delmoral, Jiahao Zhang, Jiaji Mao, Jo\~ao Manuel R.S. Tavares, Jonathan Deissler, Jun Shen, Kaouther Mouheb, Katharina Eckstein, Khadiga Omar Nassar, Klaus H. Maier-Hein, Livingstone Eli Ayivor, Marawan Elbatel, Mariam Elbakry, Matthew Choi, Maximilian Rokuss, Menna Dyab, Minghui Zhang, Mohamed Ghonim, Mohanad Ghonim, Moona Mazher, Noha Elghitany, Nouran Elghitany, Rachika E. Hamadache, Rainer Stiefelhagen, Ricardo Montoya-del-Angel, Robert Mart\'i, Salma Ahmad, Shao-Woo Yen, Stefan Klein, Steven A. Niederer, Susan Adil Ali, Toufiq Musah, Wenhui Deng, Xavier Llad\'o, Xiaodi Wang, Xiaomeng Li, Xinghua Ma, Xinyan Fang, Xin You, Xuanqi Huang, Yi Qin, Yutong Xie, Zdravko Marinov, Zhuosheng Lin.

Figure 1
Figure 1. Figure 1: TriALS cohort design and dataset characteristics. a, Number of cases per split; training bars subdivided by country (China/Egypt). b, Lesion counts per centre. Per-phase lesion counts are computed as the number of connected components within each phase’s mask. In cohorts with clustered lesions (e.g., China), contiguous coverage across contrast phases can merge adjacent non-contrast components into a single… view at source ↗
Figure 2
Figure 2. Figure 2: Inter-reader agreement between two radiologists. a, Per-case Dice violin distributions by CT phase (n = 16 per phase); annotated values above each violin are medians (venous 1.00, non-contrast 0.88, arterial 0.98, delayed 0.94). The bimodal pattern reflects cases of complete agreement (DSC ≈ 1) versus complete disagreement on lesion visibility (DSC = 0), which drives the gap between median and mean. b, Mea… view at source ↗
Figure 3
Figure 3. Figure 3: Methodological landscape of participating teams. a, Architecture backbone distribution: nnU-Net variants (ResEnc and PlainConv) dominated, with MedNeXt, xLSTM-UNet, STU-Net, and custom hybrids also represented. b, Loss function distribution: Dice + cross-entropy was used by 11 of 14 submissions; alternatives included contrastive, InfoNCE, and focal losses. 14/34 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-case Dice distributions for all participating teams and off-the-shelf baselines. Each panel shows per-case Dice pooled across all three test centres (n = 70 volumes). Transparent boxes span the interquartile range with a thick black line at the median; individual cases are overlaid as jittered points. Challenge participants are shown alongside off-the-shelf baselines VoxTell 1.1 and PASTA (orange); das… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative segmentation comparison across teams (MIC 1st, WLAI 2nd, Caladan last). a, Venous CT (External #46): all teams segment the lesion accurately (DSC 0.76–0.92). b, Same patient, NCCT: the combined GT (green) includes an occult region (yellow dashed) absent from the visible GT (blue). MIC and WLAI maintain DSC ≥ 0.90, while Caladan drops to 0.72. c, AI detects occult lesion (External #15): MIC segm… view at source ↗
Figure 6
Figure 6. Figure 6: Bootstrap ranking stability. Rankings are computed across all participating teams; only the top two (mic and wlai) are shown. 17/34 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of TriALS 2025 winner with state-of-the-art off-the-shelf methods. a, Task 1 (venous CT) per-case Dice per centre: mic_v2 (TriALS), VoxTell 1.1, and PASTA, with venous inter-reader ceiling (dashed line at 0.742). b, Task 2 combined (non-contrast) per-case Dice with NCCT inter-reader ceiling (dashed line at 0.574). c, Absolute Dice gain of mic_v2 over each off-the-shelf method per configuration. … view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison: TriALS winner vs. off-the-shelf methods. a, Egypt venous (case 101): all methods segment the large lesion, but VoxTell under-segments (DSC 0.91 vs. MIC 0.95). b, China venous (case 14): MIC achieves DSC 0.92; VoxTell produces false positives (0.73) and PASTA misses the lesion entirely (0.00). c, External NCCT (case 33): MIC maintains DSC 0.91 on non-contrast; VoxTell under-segments … view at source ↗
Figure 9
Figure 9. Figure 9: Relationship between methodological choices and performance (TriALS 2025, Egypt Task 1). a, Per-case Dice stratified by pre-training strategy. The multi-dataset pre-training approach (MIC) achieved the highest performance. b, Effect of external data usage: teams that incorporated external datasets tended to achieve higher Dice, though with notable exceptions. c, Ensemble versus single-model inference: ense… view at source ↗
Figure 10
Figure 10. Figure 10: The modality gap: cross-task degradation and annotation paradigm comparison. a, Task 1 versus Task 2 visible Dice scatter; all points fall below the diagonal, confirming non-contrast CT is harder regardless of venous-phase capability. b, Task 1 versus Task 2 combined Dice scatter. c, Combined versus visible Dice scatter per centre. d, Summary gap (∆ = visible − combined Dice) per centre; Egypt shows the l… view at source ↗
Figure 11
Figure 11. Figure 11: Cross-year progress (Egypt test set; 2024 vs. 2025). a, Task 1: best Dice improved from 0.742 to 0.791 (+4.9%). b, Task 2 combined Dice improved from 0.429 to 0.471 (+4.2%). c, Task 2 visible Dice essentially unchanged (0.55 → 0.55). d, mic team trajectory across all six settings, showing gains concentrated in the combined paradigm. 22/34 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
read the original abstract

Automated segmentation of liver lesions on non-contrast computed tomography (NCCT) is clinically important but fundamentally challenging, particularly in low-resource settings across Africa and Asia where contrast agents are frequently unavailable. Progress has been limited by the absence of annotated NCCT benchmarks. Here we describe the TriALS challenge for automated liver lesion segmentation under contrast-limited conditions, supported by a multi-centre dataset of 150 cases with four-phase CT acquisitions (600 volumes) from Egyptian and Chinese institutions. Algorithms were evaluated on 70 cases from three institutions, including an independent external cohort. The top-performing method achieved a mean venous-phase Dice of 0.754, consistent with human-level performance, yet dropped to 0.57 on NCCT. On external validation, the leading method outperformed off-the-shelf models by up to 28% in Dice on NCCT. Algorithm performance was most strongly predicted by training data scale and pre-training strategy. A cross-year comparison exposed a persistent perceptual barrier on NCCT that scaling pre-training alone cannot overcome. Data, annotations, and code are available at https://github.com/xmed-lab/TriALS.

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.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces the TriALS benchmark for automated liver lesion segmentation on non-contrast CT (NCCT) under contrast-limited conditions. It releases a multi-centre dataset of 150 cases with four-phase CT acquisitions (600 volumes) from Egyptian and Chinese institutions, evaluates participating algorithms on a 70-case test set that includes an independent external cohort, and reports that the top method reaches a mean venous-phase Dice of 0.754 (human-level) but drops to 0.57 on NCCT, with up to 28% Dice improvement over off-the-shelf models on external NCCT. Performance correlates most strongly with training data scale and pre-training strategy; a cross-year analysis indicates a persistent perceptual barrier on NCCT. Data, annotations, and code are made public.

Significance. If the empirical results hold, the work supplies a much-needed public benchmark and reproducible baseline for NCCT liver-lesion segmentation, directly addressing a clinical gap in low-resource environments. The concrete Dice scores on internal and external held-out data, together with the public release of data, annotations, and code, enable independent verification and future comparisons. The finding that scaling pre-training alone does not close the NCCT gap is a useful falsifiable observation for the field.

major comments (1)
  1. [Abstract and external-validation paragraph] Abstract and external-validation paragraph: the positioning of the benchmark as directly relevant to 'low-resource settings across Africa and Asia' rests on the untested assumption that the Egyptian-Chinese four-phase acquisitions capture scanner variability, lesion etiology distributions, and acquisition protocols typical of other low-resource sites. The reported external cohort remains within the same institutional pool and does not probe this geographic or resource-level shift, weakening the generalizability claim that underpins the paper's clinical motivation.
minor comments (2)
  1. [Methods section on data splits] Methods section on data splits: explicitly state the exact partition of the 150 cases into training, internal validation, and the 70-case test set (including how many cases belong to the external cohort) so that the 0.57 NCCT Dice and 28% gain figures can be reproduced without ambiguity.
  2. [Results tables or supplementary material] Results tables or supplementary material: report the precise off-the-shelf baseline models (e.g., nnU-Net, Swin-UNETR) and their training regimes against which the 28% Dice improvement on external NCCT is measured.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment on the generalizability of our claims. We address the point directly below and have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract and external-validation paragraph] Abstract and external-validation paragraph: the positioning of the benchmark as directly relevant to 'low-resource settings across Africa and Asia' rests on the untested assumption that the Egyptian-Chinese four-phase acquisitions capture scanner variability, lesion etiology distributions, and acquisition protocols typical of other low-resource sites. The reported external cohort remains within the same institutional pool and does not probe this geographic or resource-level shift, weakening the generalizability claim that underpins the paper's clinical motivation.

    Authors: We agree that the external cohort, while drawn from a held-out institution, remains within the same multi-center Egyptian-Chinese collection and therefore does not constitute a full geographic or resource-level shift test. The dataset does span two continents and multiple scanner vendors, supplying more diversity than typical single-center NCCT studies, yet this still falls short of validating performance across all low-resource environments. In the revised version we have moderated the abstract and introduction to describe the benchmark as addressing challenges 'particularly relevant to low-resource settings in regions such as Africa and Asia' rather than claiming direct capture of all such variability. We have also inserted a limitations paragraph that explicitly notes the need for future validation on additional sites outside the current institutional pool. revision: yes

Circularity Check

0 steps flagged

Empirical benchmark study with direct held-out measurements; no derivations or self-referential reductions.

full rationale

The paper describes a multi-centre dataset, challenge setup, and reports Dice scores from algorithm evaluations on 70 held-out cases including external validation. All central claims (0.754 venous Dice, 0.57 NCCT drop, up to 28% gain) are direct empirical measurements on split data, not derived from equations or predictions that reduce to the paper's own fitted inputs. No load-bearing self-citations, uniqueness theorems, or ansatzes are invoked in the provided text. The representativeness assumption for low-resource settings is an external generalization claim, not a circular derivation within the paper's own chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

This is an empirical benchmark and dataset paper. No free parameters are fitted to produce a central theoretical result. No new entities are postulated. Axioms are limited to standard practices in medical image analysis evaluation.

axioms (1)
  • standard math Dice coefficient is an appropriate overlap metric for evaluating lesion segmentation performance.
    Invoked implicitly in all reported performance numbers as the primary evaluation measure.

pith-pipeline@v0.9.0 · 6058 in / 1217 out tokens · 38106 ms · 2026-05-20T18:51:26.588275+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · 1 internal anchor

  1. [1]

    Reports15, 10.1038/s41598-025-02031-w (2025)

    Zhao, Y .et al.Epidemiological and demographic analysis of liver cancer attributable to modifiable risk factors from 1990 to 2021.Sci. Reports15, 10.1038/s41598-025-02031-w (2025)

    work page doi:10.1038/s41598-025-02031-w 1990

  2. [2]

    https://onlinelibrary.wiley

    El-Kassas, M.et al.Hepatocellular carcinoma in egypt in the post hcv elimination era: Changing aetiology, surveillance, and management pathways.Liver Int.46, e70595, https://doi.org/10.1111/liv.70595 (2026). https://onlinelibrary.wiley. com/doi/pdf/10.1111/liv.70595

    work page doi:10.1111/liv.70595 2026

  3. [3]

    & Akudjedu, T

    Natembeya, M., Anudjo, M., Ackah, J., Osei, M. & Akudjedu, T. The environmental sustainability implications of contrast media supply chain disruptions during the covid-19 pandemic: A document analysis of international practice guidelines. Radiography30, 43–54, 10.1016/j.radi.2024.05.017 (2024). 4.Egypt Independent. Dye shortage in Egypt for CT scans threa...

    work page doi:10.1016/j.radi.2024.05.017 2024

  4. [4]

    Masset, R

    Bilic, P.et al.The liver tumor segmentation benchmark (lits).Med. Image Analysis84, 102680, https://doi.org/10.1016/j. media.2022.102680 (2023)

    work page doi:10.1016/j 2022

  5. [5]

    The KiTS21 Challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase ct,

    Heller, N.et al.The kits21 challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary- phase ct (2023). 2307.01984

    work page arXiv 2023

  6. [6]

    (2022) AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation

    Ji, Y .et al.Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation (2022). 2206.08023

    work page arXiv 2022

  7. [7]

    Data13, 10.1038/ s41597-025-06343-4 (2025)

    Wu, X.et al.A multi-phase ct dataset for automated differential diagnosis of liver tumors.Sci. Data13, 10.1038/ s41597-025-06343-4 (2025). 9.Rokuss, M.et al.V oxtell: Free-text promptable universal 3d medical image segmentation (2025). 2511.11450. 10.Lei, W.et al.A synthetic data-driven radiology foundation model for pan-tumor clinical diagnosis (2026). 2...

    work page arXiv 2025

  8. [8]

    nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomedical Image Segmentation,

    Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J. & Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nat. Methods18, 203–211, 10.1038/s41592-020-01008-z (2020)

    work page doi:10.1038/s41592-020-01008-z 2020

  9. [9]

    MultiTalent: A Multi-dataset Approach to Medical Image Segmentation, 648–658 (Springer Nature Switzerland, 2023)

    Ulrich, C.et al. MultiTalent: A Multi-dataset Approach to Medical Image Segmentation, 648–658 (Springer Nature Switzerland, 2023)

    work page 2023

  10. [10]

    Interactive segmentation of medical images through fully convolutional neural networks

    Sakinis, T.et al.Interactive segmentation of medical images through fully convolutional neural networks.CoRR abs/1903.08205(2019). 1903.08205

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  11. [11]

    Nature Communications15(1), 654 (1 2024)

    Kirillov, A.et al.Segment anything. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 4015–4026 (2023). 15.Ma, J.et al.Segment anything in medical images.Nat. Commun.15, 10.1038/s41467-024-44824-z (2024)

    work page doi:10.1038/s41467-024-44824-z 2023

  12. [12]

    Klein, S., Staring, M., Murphy, K., Viergever, M. A. & Pluim, J. P. W. elastix: A toolbox for intensity-based medical image registration.IEEE Transactions on Med. Imaging29, 196–205, 10.1109/TMI.2009.2035616 (2010)

    work page doi:10.1109/tmi.2009.2035616 2009

  13. [13]

    Simultaneous truth and performance level estimation (STAPLE): An algorithm for the validation of image segmentation

    Warfield, S., Zou, K. & Wells, W. Simultaneous truth and performance level estimation (staple): An algorithm for the validation of image segmentation.IEEE Transactions on Med. Imaging23, 903–921, 10.1109/tmi.2004.828354 (2004)

    work page doi:10.1109/tmi.2004.828354 2004

  14. [14]

    Bias:Transparentreportingofbiomedicalimageanalysis challenges

    Maier-Hein, L.et al.Bias: Transparent reporting of biomedical image analysis challenges.Med. Image Analysis66, 101796, https://doi.org/10.1016/j.media.2020.101796 (2020). 19.Reinke, A.et al.Understanding metric-related pitfalls in image analysis validation.Nat. Methods21, 182–194, 10.1038/ s41592-023-02150-0 (2024)

    work page doi:10.1016/j.media.2020.101796 2020

  15. [15]

    Commun.9, 10.1038/s41467-018-07619-7 (2018)

    Maier-Hein, L.et al.Why rankings of biomedical image analysis competitions should be interpreted with care.Nat. Commun.9, 10.1038/s41467-018-07619-7 (2018)

    work page doi:10.1038/s41467-018-07619-7 2018

  16. [16]

    & Hochberg, Y

    Benjamini, Y . & Hochberg, Y . Controlling the false discovery rate: A practical and powerful approach to multiple testing. J. Royal Stat. Soc. Ser. B (Methodological)57, 289–300 (1995)

    work page 1995

  17. [17]

    MedNeXt: Transformer-Driven Scaling of ConvNets for Medical Image Segmentation, 405–415 (Springer Nature Switzerland, 2023)

    Roy, S.et al. MedNeXt: Transformer-Driven Scaling of ConvNets for Medical Image Segmentation, 405–415 (Springer Nature Switzerland, 2023)

    work page 2023

  18. [18]

    Huang, Z.et al.Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training.arXiv preprint arXiv:2304.06716(2023). 11/34 Train China EgyptExternal 0 20 40 60 80 Number of cases a 80 17 28 25 20 60 Cohort composition Venous NC Combined 0 50 100 150 200 250 Number of lesions b 31 45 38 222 163 260 82 ...

    work page arXiv 2023