Pith. sign in

REVIEW 3 major objections 6 minor 48 references

A frozen-encoder SAM2 adaptation plus TSDF-supervised global volume tokens raises 3D MRI segmentation from 0.58 to 0.78 mean Dice on 34 datasets, with no drop on 8 held-out datasets.

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 ↗

Freezing SAM2's encoder, fine-tuning its decoder/memory, and adding TSDF-trained global volume tokens yields 0.78 mean Dice on a new 34-dataset MRI benchmark, up from 0.58 zero-shot.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A genuinely large MRI benchmark and a sensible method paper, but the headline 'surpassing' claim rests on comparisons that are probably not like-for-like and should be fixed before this is citable as a win. the 3 major comments →

arxiv 2607.18014 v1 pith:H5UH3NIR submitted 2026-07-20 cs.CV

SAMRI-3D: Adapting SAM2 for 3D MRI Segmentation with Global Volume Tokens

Zhao Wang , Wei Dai , Hongfu Sun , Craig Engstrom , Shekhar S. Chandra This is my paper
classification cs.CV
keywords MRI segmentationSAM2 adaptation3D medical image segmentationglobal memory tokenssigned distance fieldfoundation modelsparameter-efficient fine-tuningzero-shot generalization
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.

The reading

SAMRI-3D aims to show that SAM2, a video-segmentation foundation model, can be turned into a strong general 3D MRI segmenter without retraining its image encoder. The paper reports that freezing SAM2's image encoder and fine-tuning only its decoder and memory modules lifts mean Dice from 0.58 to 0.76 on the largest MRI-only benchmark assembled to date — 10,392 volumes across 34 datasets — and that adding Global Volume Tokens trained against a truncated signed distance field (TSDF) target pushes this to 0.78 with the lowest variance across datasets. The core claim is that whole-volume boundary geometry, not appearance, is the information missing from SAM2's local streaming memory, and that a training-only TSDF reconstruction injects exactly that information. If true, this matters because it offers a parameter-efficient route to promptable MRI segmentation that generalizes to datasets never seen in training, at zero added inference cost.

Core claim

On the paper's own terms, the central claim is this: an MRI-specific adaptation of SAM2 that freezes the Hiera image encoder, fine-tunes only the lightweight decoder and memory modules, and augments the memory bank with 16 persistent Global Volume Tokens supervised by a Truncated Signed Distance Field reconstruction achieves the best mean accuracy (0.78 Dice) and lowest variance across 34 MRI datasets, surpassing recent SAM-based medical models, and — uniquely among strong baselines — shows no generalization gap on 8 unseen datasets (0.79 unseen vs 0.78 seen). The TSDF objective, discarded at inference, is what lets the tokens survive the memory attention's softmax competition and carry boun

What carries the argument

Global Volume Tokens (GVT): K=16 persistent memory tokens built by compressing the pre-computed embeddings of all slices into one whole-volume summary through Perceiver-style cross-attention, then appended to SAM2's streaming memory bank so that every slice, however far from the prompted anchor slice, sees global context. The load-bearing twist is their training objective: a training-only decoder reconstructs a Truncated Signed Distance Field of the target (boundary-weighted Smooth-L1, truncation radius 32 pixels), forcing the tokens to encode boundary geometry rather than redundant binary occupancy. The auxiliary decoder is discarded at inference, so the gain costs nothing at test time. Sli

Load-bearing premise

The reported accuracy assumes someone (or an upstream detector) supplies a reliable prompt on the cut slice — the slice with the largest target foreground — because every other slice's prediction propagates outward from that anchor; the paper does not test how accuracy degrades when the prompt lands on a slice where the target is small, fragmented, or diffuse.

What would settle it

Run SAMRI-3D on a fixed set of volumes but prompt on slices chosen progressively further from the maximum-foreground slice (for example, near the top or bottom of the target, or on a slice where the target is fragmented), holding everything else identical; if mean Dice drops steeply as the anchor moves off the cut slice, the cut-slice assumption is load-bearing. A second check: retrain GVT with a high-resolution binary mask or a different geometric prior in place of the TSDF target — the paper's own ablation already shows binary reconstruction adds little, so a positive result there would weak

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

If this is right

  • Fine-tuning only the decoder and memory modules, with the image encoder frozen, is sufficient for large gains on MRI — the paper shows SAM2 backbone size has almost no effect on zero-shot MRI accuracy, so parameter-efficient adaptation is the scalable path.
  • Whole-volume context injected through the memory bank and supervised by boundary geometry (TSDF) beats both no-token and binary-reconstruction variants on mean accuracy and variance, at under 1.1% parameter overhead and zero test-time cost.
  • The model transfers to eight datasets never seen during training with no performance drop (0.79 unseen vs 0.78 seen), which is the strongest zero-shot evidence in the comparison.
  • The improvement is concentrated exactly where it is needed: low-contrast sequences such as T1 and FLASH with diffuse boundaries, matching the paper's invisible-boundary motivation.
  • Because the TSDF decoder is discarded at inference, GVT adds no runtime cost to deployment — the reported gains are essentially free at test time.

Where Pith is reading between the lines

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

  • Not tested in the paper: sensitivity to the cut-slice anchor. The entire propagation chain starts from a prompt on the slice where the target is largest, so a natural follow-up is to sweep prompts across slices and measure how much of the 0.78 depends on that choice.
  • Beyond MRI: the recipe — persistent global tokens plus a geometric self-supervision target — is modality-agnostic and could be applied to other long-sequence inputs (CT, ultrasound, endoscopic video) where a streaming memory otherwise evicts distant context.
  • Reading the statistics: the TSDF gain over the strong baseline is statistically significant per-sample and per-label but not per-dataset (p=0.22); the paper itself positions the model as the lowest-variance, best-mean option, and that is the framing a reader should weigh — small, consistent gains rather than large jumps on any single dataset.
  • A testable prediction follows from the per-sequence analysis: on high-contrast sequences with already-sharp boundaries (DESS, T1CE) the GVT gain should shrink toward zero — which is what the paper reports, and what an independent re-run could verify.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 6 minor

Summary. The paper introduces SAMRI-3D, a benchmark and method for 3D MRI segmentation built on SAM2. The benchmark comprises 10,392 volumes from 34 datasets (26 seen, 8 unseen; 27 public, 7 in-house) across 12 anatomical domains. The method freezes SAM2.1's Hiera image encoder and prompt encoder, fine-tunes the mask decoder and memory modules, and adds Global Volume Tokens (GVT): 16 persistent tokens that summarize the whole volume and are appended to SAM2's streaming memory bank at every propagation step. The GVT tokens are trained with an auxiliary TSDF-reconstruction objective (boundary-weighted Smooth-L1 loss, Eq. 4), and the TSDF decoder is discarded at inference. The paper reports mean Dice 0.78 ± 0.14 across the 34 datasets, versus 0.58 for zero-shot SAM2-Small and 0.76 for a fine-tuned no-GVT baseline, and compares favorably with SAMed-2 (0.69), Medical-SAM2 (0.49), and SAM-Med3D (0.37), with no generalization gap on the 8 held-out datasets (0.79 unseen vs. 0.78 seen). The GVT-vs-baseline gain is significant at per-sample and per-label levels but not at the per-dataset level (p = 0.22), a result the paper reports honestly.

Significance. If the claims hold, the paper is a substantial contribution to medical foundation-model adaptation: it provides the largest MRI-only SAM2 evaluation to date with explicit seen/unseen splits; a parameter-efficient adaptation (frozen encoder, roughly 150K added trainable parameters) that approximately halves the error of zero-shot SAM2; and a falsifiable mechanistic claim — that TSDF-supervised global tokens help most on low-contrast sequences — which the per-sequence analysis (Fig. 3) directly tests. The reporting is unusually careful: per-dataset medians with quartiles, Wilcoxon tests at three aggregation levels, disclosure that the GVT advantage is not per-dataset significant, and a plan to release benchmark, code, and models. These strengths are contingent on the comparison protocol being made explicit (Major Comment 1).

major comments (3)
  1. [§4.3, Table 2; Abstract; §4.1] §4.3/Table 2: the claim that SAMRI-3D 'surpasses' SAMed-2, Medical-SAM2, and SAM-Med3D 'with strong statistical significance' is not verifiable because the evaluation protocol for these three external methods is never stated. §4.1 describes training only on the 26 seen datasets' train/validation subsets, and §4.3 labels only SAM2-Small as 'zero-shot.' If the three external methods were evaluated zero-shot with public checkpoints, the comparison is confounded by training-set access: SAMRI-3D was trained on the same datasets (hence the same anatomy, sequences, and modalities) on which it is evaluated, so the small p-values in Table S12 would show the benefit of target-distribution fine-tuning rather than architectural superiority. This is load-bearing for the headline comparative claim. Please disclose each method's checkpoint, prompt format, preprocessing, and fine-tuning status, and eith
  2. [§3.1, Fig. 1A] The method is anchored on the 'cut slice' (the slice with the largest target foreground), and segmentation propagates outward from it; at inference, boxes and/or points are supplied by the user or an upstream detector. The paper never evaluates sensitivity to prompt-slice choice. A user prompting on a slice with small or diffuse foreground may anchor the entire propagation chain poorly, and the reported 0.78 Dice is established only under the favorable cut-slice protocol. Please add an analysis that varies the anchor (e.g., slices with decreasing foreground area, or cut slices displaced by ±k slices from the optimal one) and reports the resulting Dice/HD, to demonstrate that accuracy does not critically depend on selecting the optimal slice.
  3. [§4.3] The statement that 'two further SAM2-based adaptations produced near-zero Dice on this benchmark and are omitted' is insufficient as reported. Near-zero Dice across 34 datasets is a strong signal of a protocol mismatch (different prompting requirements, evaluation convention, or preprocessing) rather than a genuine performance measurement. Name the methods, specify the protocol used (checkpoint, prompt, preprocessing, datasets on which they were run), and report the numbers; otherwise the omission invites suspicion that the comparison set was filtered.
minor comments (6)
  1. [§3.3 / Supp. B / Table S2] The abstract and §3.3 say the TSDF objective adds 'zero added cost' and that 'GVT adds no test-time cost.' Strictly, only the TSDF decoder is discarded. The GVT encoder (≈120K parameters, one extra forward pass over the pooled slice embeddings) and the appended K=16 tokens in the memory bank are active at inference, so the overhead is small but nonzero. Please rephrase and quantify the actual inference-time cost.
  2. [Table S12] The per-sample unit is not defined. With an 80/10/10 split of 10,392 volumes the test set is ≈1,039 volumes, yet n≈3,376 for the 'sample-level' tests; clarify whether the unit is volume-label pairs. Also note that samples from the same volume are not independent, so the per-dataset level (where the GVT gain is p=0.22) is the safe summary of the GVT effect.
  3. [Table 4 vs. Tables S4, S6–S8] The same model is named 'GVT-baseline' in Table 4 and 'SAMRI-3D-2dloss' in Tables S4 and S6–S8; unify the naming to avoid confusion.
  4. [Abstract; §1] 'Global V olume Tokens' contains a spurious space ('V olume') in the abstract and in the contributions list.
  5. [Fig. 3 caption] The caption does not define the 'baseline' being compared (SAMRI-3D-2dloss) nor the pooling used to compute per-sequence medians; please add these definitions.
  6. [Discussion, §5] The Discussion explicitly defers comparison with fully-supervised methods such as nnU-Net. For a paper claiming best accuracy on its own benchmark, a nnU-Net reference is important for contextualizing the absolute Dice values; if not addressed in the revision, the omission should be stated as a limitation in the conclusions.

Circularity Check

0 steps flagged

No significant circularity: the central accuracy claims are empirical, benchmarked against external baselines and controlled ablations.

full rationale

The paper's central claims are empirical results, not derivations from their own definitions. The headline improvement (0.58 → 0.76/0.78 mean Dice) is measured against zero-shot SAM2 and three independently published SAM-based medical models (SAMed-2, Medical-SAM2, SAM-Med3D), so it is externally grounded. The GVT/TSDF contribution is tested through a self-contained ablation (GVT-baseline vs. GVT-v1/v2/v3) on the same training data, and the TSDF decoder is discarded at inference, so the claim that TSDF supervision helps is not true by construction. The 8 unseen datasets provide a genuine held-out generalization check. The only self-citation, SAMRI [3], is used to justify inheriting a parameter-efficient fine-tuning strategy (freeze the encoder), not to supply evidence for the accuracy claims, so it is not load-bearing. The main caveats — possible evaluation-protocol asymmetry with prior methods and sensitivity to the cut-slice prompt choice — are methodological validity or robustness concerns, not circularity: they do not reduce any claimed result to its own input. No circular step can be exhibited from the paper's own equations or citations.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The central claims rest on standard deep-learning assumptions (pretrained SAM2, i.i.d. splits, correct labels) plus a few hand-set hyperparameters that define the GVT/TSDF mechanism. The most consequential unverified premise is the cut-slice prompting assumption. GVT is an invented component with internal, not yet independent, evidence.

free parameters (6)
  • K (number of GVT tokens) = 16
    Chosen by hand; no sweep or validation-based selection is reported (Section 3.3, Supp B).
  • TSDF truncation radius τ = 32 px
    Set by hand in Eq. (3); the boundary band width directly defines the auxiliary supervision target.
  • Boundary weighting α, σ = α=4, σ=τ/3≈10.7
    Hyperparameters of the weighted Smooth-L1 loss in Eq. (4); chosen by hand, not tuned per dataset.
  • TSDF loss weight = 1 (implicit)
    Equation (1) sums DiceFocal and TSDF losses with no explicit weight; no tuning is reported.
  • 80/20 volumetric/2D batch mix = 80% / 20%
    Training distribution choice in Section 4.2 that balances volumetric propagation with per-slice accuracy.
  • max_slices per sub-sequence (training) = 32
    Training video-cap in Supp B; volumes have 50–200 slices, so long-range context during training is truncated.
axioms (6)
  • domain assumption SAM2.1 Hiera-Small pretrained weights provide sufficiently general visual features for MRI; freezing them preserves generalization while the decoder/memory adapts to the medical domain.
    Invoked in Section 3.2 and the Discussion; the paper's entire parameter-efficient design depends on this.
  • domain assumption The slice-as-frame paradigm: treating a 3D volume as a video and propagating from the cut slice yields coherent 3D segmentations.
    Core modeling choice in Sections 2.1 and 3.1, following prior work [5, 6]; not independently validated here.
  • domain assumption The cut slice (slice with largest target foreground) is identifiable at inference and gives the best propagation anchor.
    Section 3.1 and Fig. 1A; the paper does not test sensitivity to the prompted slice position.
  • domain assumption Ground-truth labels in all 34 datasets are correct, consistently formatted, and train/test splits are independent within each dataset.
    Required for all reported Dice/HD/MSD numbers; the paper does not audit label quality.
  • domain assumption The memory attention's softmax competition is why segmentation-only GVT tokens become redundant.
    Explanatory hypothesis in Section 3.3 and Discussion; supported only indirectly by the ablations.
  • domain assumption A TSDF reconstruction objective forces tokens to encode boundary geometry complementary to appearance-based memory.
    Design hypothesis behind GVT-v3; the paper provides correlational evidence (per-sequence analysis) but no causal proof.
invented entities (1)
  • Global Volume Tokens (GVT) no independent evidence
    purpose: Persistent whole-volume memory tokens injected into SAM2's memory bank to give every slice global context; trained with a TSDF reconstruction objective discarded at inference.
    The only evidence for GVT's utility is the paper's own ablations (Table 4) and per-sequence analysis; there is no external replication or independent downstream test yet.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SAMRI-3D: Adapting SAM2 for 3D MRI Segmentation with Global Volume Tokens." pith.science (2026). https://pith.science/paper/H5UH3NIR

@misc{pith2026260718014,
  author       = {Pith},
  title        = {Pith review of: SAMRI-3D: Adapting SAM2 for 3D MRI Segmentation with Global Volume Tokens},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H5UH3NIR}},
  note         = {Machine review of arXiv:2607.18014}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Foundation models such as Segment Anything Model 2 (SAM2) have transformed natural-image and video segmentation, and recent work has begun adapting them to medical imaging. These adaptations, however, are largely general-purpose models that treat MRI as one modality among many; large-scale, MRI-specific modelling and benchmarking remain limited, even though MRI's low soft-tissue contrast leaves many boundaries effectively invisible on individual slices. We present SAMRI-3D, a benchmark and method for 3D MRI segmentation with SAM2. The SAMRI-3D benchmark is the largest MRI-only evaluation to date - 10,392 volumes from 34 datasets (27 public, 7 in-house) spanning 12 anatomical domains and 10+ sequences, with explicit seen/unseen splits. Freezing the image encoder and fine-tuning only the lightweight decoder and memory modules raises mean Dice from 0.58 (zero-shot SAM2) to 0.76, surpassing recent SAM-based medical models (SAMed-2 0.69, Medical-SAM2 0.49, SAM-Med3D 0.37) with strong statistical significance. To target invisible boundaries, we introduce Global Volume Tokens (GVT): persistent memory tokens trained with a Truncated Signed Distance Field (TSDF) reconstruction objective that is discarded at inference (zero added cost). This full model, SAMRI-3D, attains the best accuracy (0.78) and lowest variance across all 34 datasets and, uniquely, shows no drop on 8 held-out datasets (0.79 unseen vs. 0.78 seen); per-sequence analysis confirms the TSDF objective helps most where per-slice contrast is weakest. We will release the benchmark, code, and models in this paper.

Figures

Figures reproduced from arXiv: 2607.18014 by Craig Engstrom, Hongfu Sun, Shekhar S. Chandra, Wei Dai, Zhao Wang.

Figure 1
Figure 1. Figure 1: End-to-end SAMRI-3D workflow. (A) Given a 3D MRI volume and a box/point prompt on its cut slice (the slice with the largest target foreground), the model splits the volume into a forward and a backward sub-sequence and segments it slice by slice, propagating outward from the cut slice in both directions before merging the two halves into the final 3D mask. The predictor is SAM2 with a frozen image encoder … view at source ↗
Figure 2
Figure 2. Figure 2: Per-domain Dice (mean of per-dataset medians) for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-MRI-sequence improvement of SAMRI-3D over the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative segmentation results. Top rows A: representative successful cases shown in triplanar (axial, coronal, sagittal) views, [PITH_FULL_IMAGE:figures/full_fig_p007_5.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

48 extracted references · 9 linked inside Pith

  1. [1]

    Segment anything,

    A. Kirillovet al., “Segment anything,” inProc. ICCV, 2023. 1

    2023

  2. [2]

    Segment anything in medical images,

    J. Maet al., “Segment anything in medical images,”Nature Communications, vol. 15, 2024. 1

    2024

  3. [3]

    SAMRI: Segment any MRI,

    Z. Wanget al., “SAMRI: Segment any MRI,” arXiv:2510.26635, 2025. 1, 3

    Pith/arXiv arXiv 2025

  4. [4]

    SAM 2: Segment anything in images and videos,

    N. Raviet al., “SAM 2: Segment anything in images and videos,” arXiv:2408.00714, 2024. 1, 2

    Pith/arXiv arXiv 2024

  5. [5]

    Novel adaptation of video segmentation to 3D MRI: Efficient zero-shot knee segmentation with SAM2,

    A. S. Yuet al., “Novel adaptation of video segmentation to 3D MRI: Efficient zero-shot knee segmentation with SAM2,” arXiv:2408.04762, 2024. 3, 4

    Pith/arXiv arXiv 2024

  6. [6]

    Interactive 3D medical image segmentation with SAM 2,

    C. Shenet al., “Interactive 3D medical image segmentation with SAM 2,” arXiv:2408.02635, 2024. 3, 4

    Pith/arXiv arXiv 2024

  7. [7]

    SLM-SAM 2: Short-long memory SAM 2 for medical image segmentation,

    Y . Chenet al., “SLM-SAM 2: Short-long memory SAM 2 for medical image segmentation,” arXiv:2505.01854, 2025. 1, 3

    arXiv 2025

  8. [8]

    SAMed-2: Adapting SAM2 for medical image segmentation,

    Z. Chenget al., “SAMed-2: Adapting SAM2 for medical image segmentation,” inProc. MICCAI, 2025. 1, 3, 5

    2025

  9. [9]

    Segment anything in medical images and videos: Benchmark with application to echocardiography segmentation,

    J. Maet al., “Segment anything in medical images and videos: Benchmark with application to echocardiography segmentation,” arXiv, 2025. 3

    2025

  10. [10]

    RevSAM2: Reverse propagation for medical image segmentation,

    Y . Baiet al., “RevSAM2: Reverse propagation for medical image segmentation,” arXiv:2409.04298, 2024. 3

    Pith/arXiv arXiv 2024

  11. [11]

    Medical SAM 2: Segment medical images as video via segment anything model 2,

    J. Zhuet al., “Medical SAM 2: Segment medical images as video via segment anything model 2,” arXiv:2408.00874,

    Pith/arXiv arXiv

  12. [12]

    Transformer compressed sens- ing via global image tokens,

    M. Bran Lorenzanaet al., “Transformer compressed sens- ing via global image tokens,” inProc. IEEE ICIP, 2022, pp. 3011–3015. 3

    2022

  13. [13]

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

    F. Isenseeet al., “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,”Na- ture Methods, vol. 18, pp. 203–211, 2021. 1

    2021

  14. [14]

    Swin UNETR: Swin transformers for semantic segmentation of brain tumours in MRI images,

    A. Hatamizadehet al., “Swin UNETR: Swin transformers for semantic segmentation of brain tumours in MRI images,” inBrainLes, MICCAI, 2022. 1

    2022

  15. [15]

    CAN3D: Fast 3D medical image segmentation via compact context aggregation,

    W. Daiet al., “CAN3D: Fast 3D medical image segmentation via compact context aggregation,”Medical Image Analysis, vol. 82, art. 102562, 2022. 1

    2022

  16. [16]

    Semantic segmentation of 3D medical images through a kaleidoscope: Data from the Osteoarthritis Initia- tive,

    B. Wooet al., “Semantic segmentation of 3D medical images through a kaleidoscope: Data from the Osteoarthritis Initia- tive,” inProc. MIDL, PMLR, vol. 227, 2023, pp. 905–917. 1

    2023

  17. [17]

    Automated anomaly-aware 3D segmentation of bones and cartilages in knee MR images from the Os- teoarthritis Initiative,

    B. Wooet al., “Automated anomaly-aware 3D segmentation of bones and cartilages in knee MR images from the Os- teoarthritis Initiative,”Medical Image Analysis, vol. 93, art. 103089, 2024. 1

    2024

  18. [18]

    MA-SAM: Modality-agnostic SAM adap- tation for 3D medical image segmentation,

    C. Chenet al., “MA-SAM: Modality-agnostic SAM adap- tation for 3D medical image segmentation,”Medical Image Analysis, vol. 98, 2024. 1

    2024

  19. [19]

    3DSAM-adapter: Holistic adaptation of SAM from 2D to 3D for promptable medical image segmen- tation,

    S. Gonget al., “3DSAM-adapter: Holistic adaptation of SAM from 2D to 3D for promptable medical image segmen- tation,”Medical Image Analysis, vol. 98, 2024. 1

    2024

  20. [20]

    Medical SAM adapter: Adapting segment any- thing model for medical image segmentation,

    J. Wuet al., “Medical SAM adapter: Adapting segment any- thing model for medical image segmentation,”Medical Im- age Analysis, vol. 102, 2024. 1

    2024

  21. [21]

    SAM-Med3D: Towards general-purpose segmentation models for volumetric medical images,

    S. Wanget al., “SAM-Med3D: Towards general-purpose segmentation models for volumetric medical images,” arXiv,

  22. [22]

    3DMedSAM: Cross-dimensional adaptation of SAM for volumetric medical image segmenta- tion,

    A. H. Abdusalamovet al., “3DMedSAM: Cross-dimensional adaptation of SAM for volumetric medical image segmenta- tion,”Biocybernetics and Biomedical Eng., 2024. 1

    2024

  23. [23]

    SAM2-UNet: Segment anything 2 makes strong encoder for natural and medical image segmentation,

    H. Xionget al., “SAM2-UNet: Segment anything 2 makes strong encoder for natural and medical image segmentation,” arXiv:2408.08870, 2024. 1

    arXiv 2024

  24. [24]

    FocusSDF: Boundary-aware learning via signed distance supervision,

    U. Shafiqueet al., “FocusSDF: Boundary-aware learning via signed distance supervision,” arXiv:2511.11864, 2025. 3

    arXiv 2025

  25. [25]

    FlowSDF: Flow matching for med- ical image segmentation using distance transforms,

    F. Bogenspergeret al., “FlowSDF: Flow matching for med- ical image segmentation using distance transforms,”Int. J. Computer Vision, 2025. 3 8

    2025

  26. [26]

    TokenSeg: Efficient 3D medical image seg- mentation via hierarchical visual token compression,

    Z. Zenget al., “TokenSeg: Efficient 3D medical image seg- mentation via hierarchical visual token compression,”Int. J. Computer Vision, 2025. 3

    2025

  27. [27]

    Focal loss for dense object detection,

    T.-Y . Linet al., “Focal loss for dense object detection,” in Proc. ICCV, 2017. 4

    2017

  28. [28]

    Decoupled weight decay regu- larization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regu- larization,” inProc. ICLR, 2019. 4

    2019

  29. [29]

    Perceiver: General perception with iterative attention,

    A. Jaegleet al., “Perceiver: General perception with iterative attention,” inProc. ICML, 2021. 3 9 Supplementary Material A. TSDF Computation For each slice we convert the ground-truth binary mask into a Truncated Signed Distance Field (TSDF). Letd in andd out be the Euclidean distance transforms of the foreground and background. The signed distance field...

    2021

  30. [30]

    A large annotated medical image dataset for the development and evaluation of segmentation algorithms,

    A. L. Simpson et al., “A large annotated medical image dataset for the development and evaluation of segmentation algorithms,” arXiv:1902.09063, 2019; M. Antonelli et al., “The Medical Segmentation Decathlon,” Nat. Commun., 13:4128, 2022

    Pith/arXiv arXiv 1902

  31. [31]

    The multimodal brain tumor image segmentation benchmark (BRATS),

    B. H. Menze et al., “The multimodal brain tumor image segmentation benchmark (BRATS),” IEEE Trans. Med. Imag., 34(10):1993–2024, 2015; S. Bakas et al., “Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features,” Sci. Data, 4:170117, 2017; S. Bakas et al., “Segmentation labels and radiomic features for ...

    Pith/arXiv arXiv 1993

  32. [32]

    ISLES 2022: A multi-center MRI stroke lesion segmentation dataset,

    M. R. Hernandez Petzsche et al., “ISLES 2022: A multi-center MRI stroke lesion segmentation dataset,” Sci. Data, 9:762, 2022

    2022

  33. [33]

    Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH segmentation challenge,

    H. J. Kuijf et al., “Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH segmentation challenge,” IEEE Trans. Med. Imag., 38(11):2556–2568, 2019

    2019

  34. [34]

    End-to-end prostate cancer detection in bpMRI via 3D CNNs. . . ,

    A. Saha, M. Hosseinzadeh, H. Huisman, “End-to-end prostate cancer detection in bpMRI via 3D CNNs. . . ,” Med. Image Anal., 73:102155, 2021 (PI-CAI)

    2021

  35. [35]

    Variability of manual segmentation of the prostate in axial T2-weighted MRI: a multi-reader study,

    A. S. Becker et al., “Variability of manual segmentation of the prostate in axial T2-weighted MRI: a multi-reader study,” Eur. J. Radiol., 121:108716, 2019

    2019

  36. [36]

    Segment Anything in Medical Images on Laptop

    Prostate (ADC) data from the CVPR 2024 challenge “Segment Anything in Medical Images on Laptop” (MedSAM on Laptop): J. Ma, Y . Zhou, B. 27 Wang et al., “Segment Anything in Medical Images and Videos: Benchmark and Deployment,” arXiv:2408.03322, 2024; proceedings in Medical Image Segmentation Foundation Models (CVPR 2024 Challenge), LNCS 15458, Springer, 2025

    Pith/arXiv arXiv 2024

  37. [37]

    An annotated test-retest collection of prostate multiparametric MRI,

    A. Fedorov et al., “An annotated test-retest collection of prostate multiparametric MRI,” Sci. Data, 5:180281, 2018; hosted on The Cancer Imaging Archive: K. Clark et al., “The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository,” J. Digit. Imaging, 26(6):1045– 1057, 2013

    2018

  38. [38]

    Evaluation of prostate segmentation algorithms for MRI: the PROMISE12 challenge,

    G. Litjens et al., “Evaluation of prostate segmentation algorithms for MRI: the PROMISE12 challenge,” Med. Image Anal., 18(2):359–373, 2014

    2014

  39. [39]

    Automated segmentation of knee bone and cartilage combining statistical shape knowledge and CNNs: Data from the Osteoarthritis Initiative,

    F. Ambellan et al., “Automated segmentation of knee bone and cartilage combining statistical shape knowledge and CNNs: Data from the Osteoarthritis Initiative,” Med. Image Anal., 52:109–118, 2019; source imaging from the Osteoarthritis Initiative (OAI): C. G. Peterfy, E. Schneider, M. Nevitt, “The osteoarthritis initiative: report on the design rationale ...

    2019

  40. [40]

    The osteoarthritis initiative: report on the design rationale for the MRI protocol for the knee,

    Osteoarthritis Initiative (OAI): C. G. Peterfy, E. Schneider, M. Nevitt, “The osteoarthritis initiative: report on the design rationale for the MRI protocol for the knee,” Osteoarthritis Cartilage, 16(12):1433–1441, 2008; https://nda.nih.gov/oai

    2008

  41. [41]

    Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?,

    O. Bernard et al., “Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?,” IEEE Trans. Med. Imag., 37(11):2514–2525, 2018

    2018

  42. [42]

    HVSMR-2.0: A 3D cardiovascular MR dataset for whole-heart segmentation in congenital heart disease,

    D. F. Pace et al., “HVSMR-2.0: A 3D cardiovascular MR dataset for whole-heart segmentation in congenital heart disease,” Sci. Data, 11:721, 2024

    2024

  43. [43]

    TotalSegmentator MRI: robust sequence-independent segmentation of multiple anatomic structures in MRI,

    T. A. D’Antonoli et al., “TotalSegmentator MRI: robust sequence-independent segmentation of multiple anatomic structures in MRI,” Radiology, 314(2):e241613, 2025

    2025

  44. [44]

    CrossMoDA 2021 challenge: benchmark of cross-modality domain adaptation for vestibular schwannoma and cochlea segmentation,

    R. Dorent et al., “CrossMoDA 2021 challenge: benchmark of cross-modality domain adaptation for vestibular schwannoma and cochlea segmentation,” Med. Image Anal., 83:102628, 2023; original dataset: J. Shapey et al., “Segmentation of vestibular schwannoma from MRI, an open annotated dataset and baseline algorithm,” Sci. Data, 8:286, 2021

    2021

  45. [45]

    Focused shape models for hip joint segmentation in 3D magnetic resonance images,

    S. S. Chandra et al., “Focused shape models for hip joint segmentation in 3D magnetic resonance images,” Med. Image Anal., 18(3):567–578, 2014

    2014

  46. [46]

    Robust detection and segmentation for diagnosis of vertebral diseases using routine MR images,

    D. Zuki´c et al., “Robust detection and segmentation for diagnosis of vertebral diseases using routine MR images,” Comput. Graph. Forum, 33(6):190– 204, 2014

    2014

  47. [47]

    AMOS: a large-scale abdominal multi-organ benchmark for versatile medical image segmentation,

    Y . Ji et al., “AMOS: a large-scale abdominal multi-organ benchmark for versatile medical image segmentation,” Adv. Neural Inf. Process. Syst., 35:36722–36732, 2022

    2022

  48. [48]

    Tumor radiomic heterogeneity: multiparametric functional imaging to characterize variability and predict response following cervical cancer radiation therapy,

    S. R. Bowen et al., “Tumor radiomic heterogeneity: multiparametric functional imaging to characterize variability and predict response following cervical cancer radiation therapy,” J. Magn. Reson. Imaging, 47(5):1388–1396, 2018. 28

    2018

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.