Pith. sign in

REVIEW 2 major objections 4 minor 55 references

Transferability of 3D medical foundation models can be ranked without fine-tuning by comparing minimum spanning trees of features and labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 21:00 UTC pith:RE6BS3PT

load-bearing objection Solid, usable TE method for 3D medical foundation-model selection: dual MST metrics give large ranking gains and a real Bayes-error bound; the random-decoder story is only partially stress-tested but not load-bearing enough to sink the paper. the 2 major comments →

arxiv 2607.04199 v1 pith:RE6BS3PT submitted 2026-07-05 cs.CV

Topology-Driven Transferability Estimation for 3D Medical Vision Foundation Models

Jiaqi Tang , Shaoyang Zhang , Fandong Zhang , Shu Zhang , Yang Liu , Qingchao Chen This is my paper
classification cs.CV
keywords transferability estimation3D medical segmentationminimum spanning treefoundation modelstopological consistencyself-supervised learningboundary separability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

As medical vision foundation models proliferate, choosing which pre-trained encoder to fine-tune for a new 3D segmentation task becomes expensive if every candidate must be fully trained. Existing transferability scores were built for whole-image classification; they ignore voxel-wise spatial structure and boundary geometry that segmentation needs, and they often require a source-task head that self-supervised models lack. This paper shows that a sparse topological comparison is enough: build a minimum spanning tree on the dense feature embeddings and another on the semantic labels, then measure how well the two trees align. The alignment is split into a local check of whether tree edges leak across tissue boundaries and a global check of whether whole organs form cleanly separated clusters. A randomly initialized decoder is used as a free spatial projector that stabilizes the global distances. On a large multi-anatomy, multi-modality benchmark the resulting scores rank models far better than prior metrics while running roughly fifty-six times faster.

Core claim

The paper establishes that transferability of a frozen 3D medical encoder for dense segmentation can be estimated, without any fine-tuning, by the topological alignment of its feature minimum spanning tree with the corresponding semantic-label minimum spanning tree, measured at both local boundary and global anatomical scales and fused by a task-complexity gate.

What carries the argument

Dual MST metrics—Local Boundary-Aware Topological Consistency (LBTC), whose leakage rate is proved to lower-bound the Bayes error, and Global Representation Topology Divergence (GRTD)—together with the claim that a randomly initialized decoder acts as a variance-reducing, topology-preserving spatial projector.

Load-bearing premise

The claim that a randomly initialized convolutional decoder with upsampling preserves the feature MST edge set while strictly reducing the variance of pairwise distances, so that global topology can safely be read from decoder features.

What would settle it

Replace the random nnU-Net-style decoder with a pure attention or diffusion decoder of matched capacity; if GRTD computed on those features loses rank correlation with fine-tuning Dice while LBTC remains unchanged, the spatial-projector justification fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Model zoos can be ranked for a new 3D segmentation task in seconds rather than thousands of GPU-minutes of exhaustive fine-tuning.
  • Local leakage rate supplies a training-free, finite-sample certificate of boundary separability that is theoretically linked to Bayes error.
  • Task complexity (number of classes) automatically re-weights local versus global topology, so the same scorer works for both focal lesions and multi-organ layouts.
  • The same MST machinery extends, via a topological label-coherence rate, to classification and regression transferability estimation.

Where Pith is reading between the lines

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

  • Layer-wise LBTC and GRTD profiles could be turned into a diagnostic that decides which layers to freeze or adapt during parameter-efficient fine-tuning.
  • Periodic MST scoring on a small labeled probe set during pre-training could serve as an early-stopping signal for transferability rather than reconstruction loss alone.
  • If the random-decoder projection generalizes, the same variance-reduction argument may justify using untrained upsamplers as free topological regularizers in other dense-prediction settings.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a training-free transferability estimation (TE) method for selecting among 3D medical vision foundation models for dense segmentation. It constructs Minimum Spanning Trees on dense features and on semantic labels, then measures alignment at two scales: Local Boundary-Aware Topological Consistency (LBTC), whose MST leakage rate is proved to lower-bound the Bayes error (Theorem 3.2), and Global Representation Topology Divergence (GRTD), which compares tree-weight discrepancies after a randomly initialized decoder. The decoder is justified as a topology-preserving spatial projector via a Johnson–Lindenstrauss-style MST preservation argument (Theorem 3.3) and a variance-reduction claim for upsampling (Proposition 3.4). The two scores are fused by a task-adaptive gate α = tanh(log|C|/2). On the OpenMind zoo (114 k volumes, 7 SSL encoders, 5 ID + 2 OOD segmentation tasks) the method reports average weighted Kendall τ = 0.638 (0.36 above CCFV) and a 56× speed-up over that baseline.

Significance. If the ranking gains hold under the requested ablations, the work supplies the first practical, non-parametric TE tool tailored to SSL-based 3D medical foundation models for dense prediction—an important and previously underserved setting. Strengths that should be credited include: (i) explicit finite-sample and asymptotic bounds linking MST leakage to 1-NN and Bayes error (Appendix E.1), (ii) a clean JL-style argument for MST edge-set preservation under random projection, (iii) a large multi-task ID/OOD benchmark with clear gains over classification-oriented TE metrics and the only prior segmentation TE baseline (CCFV), and (iv) substantial wall-clock acceleration. These elements make the contribution both theoretically grounded and immediately usable for model selection.

major comments (2)
  1. [§3.3, Table 2 (right)] §3.3 and Table 2 (right): The claim that a randomly initialized decoder is necessary for stable GRTD is load-bearing for the dual-scale design and for the theoretical narrative, yet the only empirical support is Effective Rank / Participation Ratio expansion under NoSkip. There is no ablation of weighted Kendall τ when GRTD is computed on encoder features versus decoder features (with and without skips). Without that ranking ablation the necessity of the decoder step—and therefore the practical force of Theorems 3.3 and Proposition 3.4—remains untested. A short table reporting τ for GRTD-encoder vs. GRTD-decoder on the same tasks would resolve the issue.
  2. [§4.1, Appendix B] §4.1 / Appendix B: Free parameters (foreground/background budgets 256/2560, Gaussian layer centers/sigmas, LBTC N and w) are fixed without a systematic sensitivity study that shows ranking stability under reasonable variation. Appendix D reports limited seed and initialization robustness and some OAT checks for LBTC, but the main claim of “parameter-free” adaptive fusion still rests on these choices. Reporting τ under a small grid of budgets and (σd, σe) would strengthen confidence that the SOTA margin is not an artifact of the chosen defaults.
minor comments (4)
  1. [Abstract] Abstract and Table 1: the abstract states “average weighted Kendall (outperforming by 0.36)” without naming the coefficient; Table 1 reports 0.638 vs. CCFV 0.272, which matches, but the abstract should explicitly write “weighted Kendall’s τ”.
  2. [Figure 3, Eq. (2)] Figure 3 caption and Eq. (2): the notation for δc mixes summation over Ec of feature weights with semantic weights; a short clarifying sentence would help readers who do not immediately see that wsem is zero inside classes.
  3. [§2.2] §2.2: CCFV is the only segmentation-specific baseline; a one-sentence note on why other recent geometric TE methods (e.g., OTCE, ETran) were not adapted would be useful for completeness.
  4. [§5] Limitations: the paper correctly flags that the random-decoder argument is specific to convolution-style upsampling; this caveat should also appear briefly in the main-text discussion of GRTD so that readers do not over-generalize.

Circularity Check

0 steps flagged

No circularity: dual MST metrics, Bayes/JL bounds, and task-adaptive fusion are independently defined and scored against held-out fine-tuning Dice; no fitted inputs or self-referential reductions.

full rationale

The paper's central claims rest on non-parametric constructions (feature MST vs. semantic MST) whose scores (LBTC leakage rate, GRTD tree-weight discrepancy) are computed solely from frozen encoder features, a randomly initialized decoder, and target labels, then ranked against independent fine-tuning Dice obtained after full training. Theorem 3.2 derives the finite-sample lower bound ho_n ≥ n/(2(n-1)) R_1NN_n (and liminf ho ≥ R*/2) from the classical cut property of MSTs, 1-NN edge inclusion, and the Cover-Hart sandwich; the proof never inserts the empirical ranking results. Theorems 3.3 and Proposition 3.4 likewise invoke only the Johnson-Lindenstrauss lemma and chi-squared variance of random projections/upsampling, again without reference to the downstream Kendall-τ numbers. The gating α = tanh(log|C|/2) is a fixed, parameter-free function of semantic cardinality, not a fit to performance. No self-citation supplies a uniqueness theorem or ansatz that forces the metric; the OpenMind zoo is an external benchmark. Consequently the derivation chain is self-contained and the reported SOTA correlations are genuine out-of-sample evaluations rather than tautologies.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The method rests on classical graph and concentration results plus a handful of engineering choices (sampling budgets, Gaussian layer weights, tanh gate) that are not derived from first principles. No new physical entities are postulated; the invented constructs are purely algorithmic.

free parameters (4)
  • foreground/background sampling budgets = 256 / 2560
    Fixed at 256 fg + 2560 bg voxels per volume; shown to be rank-stable but still a free design choice that affects compute and variance.
  • Gaussian layer-weight centers and sigmas = (0.8,0.9)
    μd=2, σd=0.8 for decoder; μe=3, σe=0.9 for encoder; chosen by hand and only partially ablated.
  • task-adaptive gate α = tanh(log|C|/2) = tanh(κ/2)
    Functional form and the factor 1/2 are design choices; fixed-α ablations show sensitivity.
  • LBTC boundary-patch count N and patch size w = N=50, w=16
    Base values N=50, w=16; sensitivity tables provided but values remain free.
axioms (4)
  • domain assumption Feature–label alignment principle: transferability is inversely related to Bayes error of the frozen encoder features (Assumption 3.1).
    Standard in TE literature; invoked to justify using leakage as a proxy.
  • standard math All pairwise distances are almost surely distinct under continuous marginals, so the MST is unique (used in Thm 3.2).
    Classical measure-theoretic fact.
  • standard math Random convolutions + upsampling realize a Johnson–Lindenstrauss-style projection that preserves MST edges with high probability (Thm 3.3).
    JL lemma + cut property; the modeling of transposed-convolution kernels as independent Gaussian slices is an additional modeling step.
  • ad hoc to paper Semantic graph edge weights can be set to zero inside classes and to a median-distance penalty across classes.
    Convenient construction that makes W(Tsem) a pure inter-class cost; not forced by theory.
invented entities (3)
  • Local Boundary-Aware Topological Consistency (LBTC) independent evidence
    purpose: Quantifies fraction of MST edges that cross class boundaries inside local patches around morphological gradients.
    New metric; independent evidence is the finite-sample Bayes lower bound proved in Thm 3.2.
  • Global Representation Topology Divergence (GRTD) no independent evidence
    purpose: Normalized per-class tree-weight discrepancy between feature MST and semantic MST.
    New metric; justified by the random-decoder projector argument.
  • Task-adaptive topological fusion gate no independent evidence
    purpose: Maps log(number of classes) through tanh to weight GRTD vs LBTC.
    Heuristic that improves average ranking; no external falsifiable prediction.

pith-pipeline@v1.1.0-grok45 · 26809 in / 2944 out tokens · 38682 ms · 2026-07-11T21:00:06.755170+00:00 · methodology

0 comments
read the original abstract

The growing number of medical vision foundation models highlights the need for effective model selection. However, mainstream selection methods rely on exhaustive fine-tuning, which is computationally expensive. Most of the existing Transferability Estimation (TE) metrics are primarily designed for image-level classification. They fail to preserve spatial relationships and fine-grained boundary details, which are crucial for the segmentation task. Additionally, while image-level tasks typically process a single feature vector per input, dense prediction tasks in 3D medical imaging require voxel-wise evaluation against dense annotations. To bridge these gaps, we propose a \textit{non-parametric, topology-driven} framework that estimates transferability directly from the alignment between the sparse 1-skeleton graph of dense features and semantic labels via Minimum Spanning Trees (MST). We decouple the alignment into two complementary geometric scales: Local Boundary-Aware Topological Consistency (LBTC) to assess boundary separability, where we prove that the MST leakage rate serves as a finite-sample lower bound on the Bayes error; and Global Representation Topology Divergence (GRTD) to evaluate the overall anatomical layout. Crucially, we formally justify a counterintuitive mechanism: Although without fine-tuning, the randomly initialized segmentation decoder acts as a topology-preserving spatial projector, reducing the variance of pairwise distance estimates and stabilizing global alignment evaluation. Fused via a task-adaptive gating mechanism, these dual metrics adapt to diverse clinical complexities. Evaluated on a large-scale benchmark of 114,000 3D medical volumes across diverse anatomical tasks, our topological framework achieves state-of-the-art transferability estimation with an average weighted Kendall (outperforming by 0.36) while accelerating evaluation by 56 times.

Figures

Figures reproduced from arXiv: 2607.04199 by Fandong Zhang, Jiaqi Tang, Qingchao Chen, Shaoyang Zhang, Shu Zhang, Yang Liu.

Figure 1
Figure 1. Figure 1: Motivation and core contributions of our [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: The radar chart evaluates models on datasets with diverse anatomical regions and modalities: including COS [5] (carotid vessel, MRI), SBM [13] (brain, MRI), AMO [25] (abdominal organs, CT/MRI), KIT [18] (kidney, CT), HAN [35] (head and neck, MRI), and TPC [47] (brain vessels, MRI). Fine-tuning performance of diverse foundation models from [43] varies across downstream datasets. Right: Previous method… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our Topology-Driven TE framework. By constructing MST, the framework [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Correlation between the fine-tuning performance and transferability metrics using MSF as [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

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

55 extracted references · 1 canonical work pages

  1. [1]

    An information-theoretic approach to transferability in task transfer learning

    Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In2019 IEEE international conference on image processing (ICIP), pages 2309–2313. IEEE, 2019

    2019

  2. [2]

    Cervénanský, et al

    Olivier Bernard, Alain Lalande, Clement Zotti, J. Cervénanský, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: Is the problem solved? IEEE Transactions on Medical Imaging, 2018

    2018

  3. [3]

    Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet.PLoS medicine, 15(11):e1002699, 2018

    Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet.PLoS medicine, 15(11):e1002699, 2018

    2018

  4. [4]

    Scalable diverse model selection for accessible transfer learning.Advances in Neural Information Processing Systems, 34:19301– 19312, 2021

    Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning.Advances in Neural Information Processing Systems, 34:19301– 19312, 2021

    2021

  5. [5]

    Huijun Chen, Xihai Zhao, Jiaqi Dou, C Du, R Yang, H Sun, S Yu, H Zhao, C Yuan, and N Balu. Carotid vessel wall segmentation and atherosclerosis diagnosis challenge.Runyu Yang, Haozhong Sun, Shuwan Yu, Huilin Zhao, Chun Yuan, and Niranjan Balu,“Carotid vessel wall segmentation and atherosclerosis diagnosis challenge, 2022

    2022

  6. [6]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PmLR, 2020

    2020

  7. [7]

    Masked image modeling advances 3d medical image analysis

    Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970–1980, 2023

    1970

  8. [8]

    The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism.Molecular psychiatry, 19(6):659–667, 2014

    Adriana Di Martino, Chao-Gan Yan, Qingyang Li, Erin Denio, Francisco X Castellanos, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Susan Y Bookheimer, Mirella Dapretto, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism.Molecular psychiatry, 19(6):659–667, 2014

    2014

  9. [9]

    Pactran: Pac- bayesian metrics for estimating the transferability of pretrained models to classification tasks

    Nan Ding, Xi Chen, Tomer Levinboim, Soravit Changpinyo, and Radu Soricut. Pactran: Pac- bayesian metrics for estimating the transferability of pretrained models to classification tasks. InEuropean Conference on Computer Vision, pages 252–268. Springer, 2022

    2022

  10. [10]

    Phikon-v2, a large and public feature extractor for biomarker prediction.arXiv preprint arXiv:2409.09173, 2024

    Alexandre Filiot, Paul Jacob, Alice Mac Kain, and Charlie Saillard. Phikon-v2, a large and public feature extractor for biomarker prediction.arXiv preprint arXiv:2409.09173, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Etran: Energy-based transferability estimation

    Mohsen Gholami, Mohammad Akbari, Xinglu Wang, Behnam Kamranian, and Yong Zhang. Etran: Energy-based transferability estimation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 18613–18622, 2023

    2023

  12. [12]

    Understanding the difficulty of training deep feedfor- ward neural networks

    Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedfor- ward neural networks. InProceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010

    2010

  13. [13]

    Deep learning enables automatic detection and segmentation of brain metastases on multisequence mri.Journal of Magnetic Resonance Imaging, 51(1):175–182, 2020

    Endre Grøvik, Darvin Yi, Michael Iv, Elizabeth Tong, Daniel Rubin, and Greg Zaharchuk. Deep learning enables automatic detection and segmentation of brain metastases on multisequence mri.Journal of Magnetic Resonance Imaging, 51(1):175–182, 2020

    2020

  14. [14]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop, pages 272–284. Springer, 2021

    2021

  15. [15]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000–16009, June 2022. 10

    2022

  16. [16]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015

    2015

  17. [17]

    Vista3d: A unified segmentation foundation model for 3d medical imaging

    Yufan He, Pengfei Guo, Yucheng Tang, Andriy Myronenko, Vishwesh Nath, Ziyue Xu, Dong Yang, Can Zhao, Benjamin Simon, Mason Belue, et al. Vista3d: A unified segmentation foundation model for 3d medical imaging. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 20863–20873, 2025

    2025

  18. [18]

    The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge

    Nicholas Heller, Fabian Isensee, Klaus H Maier-Hein, Xiaoshuai Hou, Chunmei Xie, Fengyi Li, Yang Nan, Guangrui Mu, Zhiyong Lin, Miofei Han, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical image analysis, 67:101821, 2021

    2021

  19. [19]

    Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset

    Moritz R Hernandez Petzsche, Ezequiel De La Rosa, Uta Hanning, Roland Wiest, Waldo Valenzuela, Mauricio Reyes, Maria Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, et al. Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data, 9(1):762, 2022

    2022

  20. [20]

    Asymptotic theory of greedy approximations to minimal k-point random graphs.IEEE Transactions on Information Theory, 45(6):1921–1938, 1999

    Alfred O Hero and Olivier JJ Michel. Asymptotic theory of greedy approximations to minimal k-point random graphs.IEEE Transactions on Information Theory, 45(6):1921–1938, 1999

    1921

  21. [21]

    Green hierarchical vision transformer for masked image modeling.Advances in Neural Information Processing Systems, 35:19997–20010, 2022

    Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling.Advances in Neural Information Processing Systems, 35:19997–20010, 2022

    2022

  22. [22]

    Frustratingly easy transferability estimation

    Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. InInternational conference on machine learning, pages 9201–9225. PMLR, 2022

    2022

  23. [23]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods, 18(2):203–211, 2021

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods, 18(2):203–211, 2021

    2021

  24. [24]

    nnu-net revisited: A call for rigorous validation in 3d medical image segmentation

    Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, pages 488–498. Springer, 2024

    2024

  25. [25]

    Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation.Advances in neural information processing systems, 35:36722–36732, 2022

    Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation.Advances in neural information processing systems, 35:36722–36732, 2022

    2022

  26. [26]

    On dataset transferability in medical image classification.arXiv preprint arXiv:2412.20172, 2024

    Dovile Juodelyte, Enzo Ferrante, Yucheng Lu, Prabhant Singh, Joaquin Vanschoren, and Veronika Cheplygina. On dataset transferability in medical image classification.arXiv preprint arXiv:2412.20172, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Peftdiff: Diffusion-guided transferability estimation for parameter-efficient fine-tuning

    Prafful Kumar Khoba, Zijian Wang, Chetan Arora, and Mahsa Baktashmotlagh. Peftdiff: Diffusion-guided transferability estimation for parameter-efficient fine-tuning. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1454–1463, 2025

    2025

  28. [28]

    Ranking neural checkpoints

    Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2663–2673, 2021

    2021

  29. [29]

    On variants of the johnson–lindenstrauss lemma.Random Structures & Algo- rithms, 33(2):142–156, 2008

    Jiˇrí Matoušek. On variants of the johnson–lindenstrauss lemma.Random Structures & Algo- rithms, 33(2):142–156, 2008

    2008

  30. [30]

    Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching.Advances in Neural Information Processing Systems, 36:27922–27950, 2023. 11

    2023

  31. [31]

    Muslim, Syamsiah Mashohor, Gheyath Al Gawwam, Rozi Mahmud, Marsyita Binti Hanafi, Osama Al-nuaimi, Raad Josephine, and Abdullah Dhaifallah Almutairi

    Ali M. Muslim, Syamsiah Mashohor, Gheyath Al Gawwam, Rozi Mahmud, Marsyita Binti Hanafi, Osama Al-nuaimi, Raad Josephine, and Abdullah Dhaifallah Almutairi. Brain mri dataset of multiple sclerosis with consensus manual lesion segmentation and patient meta information.Data in Brief, 42:108139, 2022

    2022

  32. [32]

    Leep: A new measure to evaluate transferability of learned representations

    Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. InInternational Conference on Machine Learning, pages 7294–7305. PMLR, 2020

    2020

  33. [33]

    Trans- ferability estimation using bhattacharyya class separability

    Michal Pándy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Trans- ferability estimation using bhattacharyya class separability. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9172–9182, 2022

    2022

  34. [34]

    Exploring scalable medical image encoders beyond text supervision.Nature Machine Intelligence, 7(1):119–130, 2025

    Fernando Pérez-García, Harshita Sharma, Sam Bond-Taylor, Kenza Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Exploring scalable medical image encoders beyond text supervision.Nature Machine Intelligence, 7(1):119–130, 2025

    2025

  35. [35]

    Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset.Medical physics, 50(3):1917–1927, 2023

    Gašper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset.Medical physics, 50(3):1917–1927, 2023

    1917

  36. [36]

    A large open access dataset of brain metastasis 3d segmentations on mri with clinical and imaging information.Scientific Data, 11(1):254, 2024

    Divya Ramakrishnan, Leon Jekel, Saahil Chadha, Anastasia Janas, Harrison Moy, Nazanin Maleki, Matthew Sala, Manpreet Kaur, Gabriel Cassinelli Petersen, Sara Merkaj, et al. A large open access dataset of brain metastasis 3d segmentations on mri with clinical and imaging information.Scientific Data, 11(1):254, 2024

    2024

  37. [37]

    Oriane Siméoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Otce: A transferability metric for cross-domain cross-task representations

    Yang Tan, Yang Li, and Shao-Lun Huang. Otce: A transferability metric for cross-domain cross-task representations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15779–15788, 2021

    2021

  39. [39]

    Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh

    Yucheng Tang, Dong Yang, Wenqi Li, Holger R. Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20730–20740, June 2022

    2022

  40. [40]

    Transferability and hardness of supervised classification tasks

    Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. InProceedings of the IEEE/CVF international conference on computer vision, pages 1395–1405, 2019

    2019

  41. [41]

    A weighted correlation index for rankings with ties

    Sebastiano Vigna. A weighted correlation index for rankings with ties. InProceedings of the 24th international conference on World Wide Web, pages 1166–1176, 2015

    2015

  42. [42]

    Training dataset for hntsmrg 2024 challenge.https://zenodo.org/doi/10.5281/zenodo.11199559, 2024

    Kareem Wahid, Cem Dede, Mohamed Naser, and Clifton Fuller. Training dataset for hntsmrg 2024 challenge.https://zenodo.org/doi/10.5281/zenodo.11199559, 2024

    work page doi:10.5281/zenodo.11199559 2024

  43. [43]

    An openmind for 3d medical vision self- supervised learning

    Tassilo Wald, Constantin Ulrich, Jonathan Suprijadi, Sebastian Ziegler, Michal Nohel, Robin Peretzke, Gregor Kohler, and Klaus Maier-Hein. An openmind for 3d medical vision self- supervised learning. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 23839–23879, 2025

    2025

  44. [44]

    Triad: Vision foundation model for 3d magnetic resonance imaging.arXiv preprint arXiv:2502.14064, 2025

    Shansong Wang, Mojtaba Safari, Qiang Li, Chih-Wei Chang, Richard LJ Qiu, Justin Roper, David S Yu, and Xiaofeng Yang. Triad: Vision foundation model for 3d magnetic resonance imaging.arXiv preprint arXiv:2502.14064, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    V oco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis

    Linshan Wu, Jiaxin Zhuang, and Hao Chen. V oco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22873–22882, 2024. 12

    2024

  46. [46]

    Delving into masked autoencoders for multi-label thorax disease classification

    Junfei Xiao, Yutong Bai, Alan Yuille, and Zongwei Zhou. Delving into masked autoencoders for multi-label thorax disease classification. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3588–3600, 2023

    2023

  47. [47]

    Topcow: Benchmarking topology-aware anatomical segmentation of the circle of willis (cow) for cta and mra.Journal of Medical Imaging Technology, 34:123–135, 2024

    K Yang, F Musio, Y Ma, and N Juchler. Topcow: Benchmarking topology-aware anatomical segmentation of the circle of willis (cow) for cta and mra.Journal of Medical Imaging Technology, 34:123–135, 2024

    2024

  48. [48]

    Pick the best pre-trained model: Towards transferability estimation for medical image segmentation

    Yuncheng Yang, Meng Wei, Junjun He, Jie Yang, Jin Ye, and Yun Gu. Pick the best pre-trained model: Towards transferability estimation for medical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 674–683. Springer, 2023

    2023

  49. [49]

    Logme: Practical assessment of pre-trained models for transfer learning

    Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. InInternational Conference on Machine Learning, pages 12133–12143. PMLR, 2021

    2021

  50. [50]

    Omnirad: A radiological foundation model for multi-task medical image analysis.arXiv preprint arXiv:2602.04547, 2026

    Luca Zedda, Andrea Loddo, and Cecilia Di Ruberto. Omnirad: A radiological foundation model for multi-task medical image analysis.arXiv preprint arXiv:2602.04547, 2026

    arXiv 2026

  51. [51]

    Models genesis.Medical image analysis, 67:101840, 2021

    Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis.Medical image analysis, 67:101840, 2021. A The Foundation Model Zoo and Downstream Tasks We utilize the extensive library of pre-trained models provided by the OpenMind benchmark [43]. In contrast to prior transferability evaluation studies that used models trai...

    2021

  52. [52]

    , C seg}Z′Y ′X ′ , where −1 denotes the padded region and is excluded from all computations

    Label downsampling.The voxel-level segmentation mask is nearest-neighbour in- terpolated to the feature-map resolution to obtain a per-voxel label array y∈ {−1,0,1, . . . , C seg}Z′Y ′X ′ , where −1 denotes the padded region and is excluded from all computations

  53. [53]

    Exactly256 foreground voxelsare drawn uniformly at random per case, stratified across foreground classes so that each class receives an approximately equal quota

    Foreground sampling.V oxels with label ≥1 are treated as foreground. Exactly256 foreground voxelsare drawn uniformly at random per case, stratified across foreground classes so that each class receives an approximately equal quota

  54. [54]

    For metrics that require class-distribution estimation (GBC, LogME, GRTD), background samples are retained and treated as class c= 0

    Background sampling.V oxels with label= 0 are subsampled to at most2560 background voxelsper case. For metrics that require class-distribution estimation (GBC, LogME, GRTD), background samples are retained and treated as class c= 0 . For LEEP, background samples are excluded (bg_sample_num= 0) since background voxels carry no semantic label. 14

  55. [55]

    Green AI

    Global (Fv) sampling.A separate global pool of voxels (foreground and background jointly, up to fv_sample_num) is maintained per case for the Feature Variety component of CCFV . Experiments compute resourcesAll experiments were conducted on a workstation equipped with a single NVIDIA RTX 5090 GPU (32GB VRAM), an Intel® Xeon® Platinum 8470Q CPU (25 vCPUs),...

    arXiv 2090