Pith. sign in

REVIEW 3 major objections 6 minor 35 references

BATS keeps 3D medical segmentation nearly as accurate as dense models while cutting peak GPU memory by more than half, by spending fine resolution only near predicted class boundaries.

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

2026-07-30 19:50 UTC pith:6RB3NNVM

load-bearing objection Solid 3D systems paper: near MedNeXt-L Dice with >53% peak-memory cut under nnU-Net Revisited; runtime tracks boundary density and the occupancy oracle is the real soft spot, not a hidden collapse. the 3 major comments →

arxiv 2607.26829 v1 pith:6RB3NNVM submitted 2026-07-29 cs.CV cs.LG

BATS: Resource-Efficient Volumetric Segmentation with Boundary-Aware Mixed-Resolution Tokens

David Hagerman , Roman Naeem , Fredrik Kahl This is my paper
classification cs.CV cs.LG
keywords 3D medical image segmentationboundary-aware tokensmixed-resolution hierarchyparent cluster attentionmemory-efficient inferenceadaptive token selectionvolumetric CT/MRI
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.

Dense 3D segmentation networks treat every voxel the same, so activation memory and runtime grow with the full volume even where tissue is homogeneous. BATS instead builds an input-dependent mixed-resolution token hierarchy: a dense boundary predictor scores every split level in parallel, a fine-first cascade keeps selected fine tokens and their ancestors, and a sparse refiner processes only that hierarchy before rasterising back to a dense mask. Parent cluster attention injects each token's coarser ancestors into its local neighbourhood so cross-scale context does not require dense multi-scale feature maps. On five public CT/MRI benchmarks under a fixed protocol, BATS sits within 0.37 Dice of the strongest dense baseline of similar size, posts the best liver-tumour score, and uses less than half the peak allocated GPU memory on the three datasets measured. Runtime falls when the retained hierarchy stays sparse and rises when boundaries fill the crop, so the method's practical win is consistent memory relief with a dataset-dependent speed trade-off.

Core claim

Boundary-aware mixed-resolution tokens can replace dense multi-scale feature maps in volumetric medical segmentation: by predicting split relevance independently at every scale, cascading from fine to coarse so no coarse miss permanently erases fine evidence, and refining the resulting sparse hierarchy with parent cluster attention, BATS matches dense accuracy within a fraction of a Dice point while cutting peak allocated GPU memory by more than 53 percent on KiTS, LiTS, and BraTS.

What carries the argument

The fine-first context cascade plus parent cluster attention: parallel dense split scores at four patch scales are turned into a variable mixed-resolution hierarchy that always retains ancestor chains, then each token's local attention neighbourhood is augmented with its hierarchical ancestors so cross-scale context is deterministic and sparse.

Load-bearing premise

That a patch needing fine resolution is essentially one that already contains more than one semantic class, so binary multi-class occupancy is a stable enough signal to decide where the expensive tokens go.

What would settle it

A dataset or ablation in which clinically critical thin or low-contrast structures are routinely missed by the multi-class-occupancy split targets, so that raising recall of the boundary heads still fails to recover Dice while token count (and thus memory/runtime) stays high.

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

If this is right

  • Peak inference memory for high-resolution 3D crops can be halved without switching to a weaker architecture family.
  • Runtime gains appear mainly on scans whose predicted boundaries occupy a small fraction of the volume; dense-boundary tasks pay a token overhead.
  • Parallel multi-scale boundary scoring plus ancestor retention removes the irreversible miss path of iterative coarse-to-fine token gating.
  • Cross-scale context can be supplied by deterministic parent links instead of dense feature pyramids or nearest-neighbour search over tokens.

Where Pith is reading between the lines

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

  • The same recall-oriented cascade could be applied to other sparse 3D tasks (detection, registration) where missing a small structure is costlier than over-retaining tokens.
  • Further memory wins likely require sparsifying the still-dense boundary predictor and segmentation head, which the paper flags as open.
  • Boundary density of a dataset becomes a first-class systems variable: protocol designers may want to report retained-token statistics alongside Dice.

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. BATS proposes a 3D volumetric segmentation architecture that replaces dense multi-scale feature maps with an input-dependent mixed-resolution token hierarchy. A dense ConvNeXt–FPN boundary predictor scores split decisions in parallel at four patch scales; a fine-first context cascade retains selected fine tokens and their ancestor chains; a refiner with parent cluster attention injects hierarchical ancestors into local attention neighbourhoods; and a lightweight head rasterises the sparse hierarchy to dense logits. Under the standardised nnU-Net Revisited protocol on ACDC, LiTS, BraTS, KiTS, and AMOS, BATS reports the highest LiTS Dice (83.39), averages within 0.37 Dice of MedNeXt-L, cuts peak allocated GPU memory by more than 53% on KiTS/LiTS/BraTS, and shows runtime gains on sparse hierarchies (KiTS/LiTS) but a slowdown on dense BraTS. Ablations compare parallel vs iterative vs random hierarchy construction, cross-scale attention variants, and component removals.

Significance. Activation memory and inference cost are genuine bottlenecks for high-resolution 3D medical segmentation. The paper’s contribution is a complete, carefully engineered pipeline—parallel multi-scale boundary scoring, recall-oriented fine-first cascade, and parent cluster attention—that converts predicted class transitions into a variable-size mixed-resolution hierarchy without dense multi-scale refiner maps. Strengths include evaluation under the nnU-Net Revisited protocol with matched folds and baselines, matched-setting peak-memory and runtime measurements on A100, and ablations that isolate hierarchy construction (including split-depth precision/recall) and cross-scale interaction. The honest reporting that runtime tracks retained-token density, including a BraTS slowdown, strengthens rather than weakens the efficiency claim. If the results hold under broader hardware and task scrutiny, BATS is a credible memory-efficient alternative to dense ConvNet baselines when boundary structure is sparse.

major comments (3)
  1. [§3.2, Table 5, Appendix C] §3.2 defines the split target as binary multi-class occupancy (max_patch(labels) ≠ min_patch(labels)) with τ=0.5. Table 5 reports coarse-scale precision of only 73.8% for BATS versus 93.6% for the iterative baseline, and Fig. 3 / Appendix C show fine tokens on non-semantic intensity edges and irreversible misses when boundary recall fails. This proxy is load-bearing for both the accuracy floor and the claim that runtime follows boundary density (Tables 3–4; BraTS ~5× denser and slower). The manuscript should quantify how often segmentation errors coincide with under-resolved regions versus classification failures under adequate resolution (Appendix C is qualitative only), and state more explicitly for which target geometries (thin vessels, low-contrast lesions, texture-defined classes) the occupancy oracle is expected to be insufficient.
  2. [§4.3, Table 5] The central architectural claim that parallel scoring plus fine-first cascade outperforms iterative coarse-to-fine gating rests on Table 5 (87.23 vs 84.55 HEC Dice) and the associated split-depth recall gap, but this comparison is reported on a single KiTS fold at matched boundary-predictor capacity. Because the iterative baseline also swaps the dense CNN for attention-based blocks, capacity matching does not fully isolate parallel vs sequential selection. Multi-fold or multi-dataset confirmation (at least KiTS + one denser set such as BraTS or AMOS) is needed before the parallel/fine-first design can be treated as established rather than suggestive.
  3. [§4.2, Table 4, Table 3] Efficiency (Table 4) is measured only against MedNeXt-L k3/k5 on KiTS, LiTS, and BraTS. AMOS retains 33.9k tokens/crop (Table 3)—between LiTS and BraTS—yet has no memory/runtime numbers; ACDC is likewise absent. Appendix D acknowledges the limited comparator set. Given that the abstract’s efficiency headline is “consistent memory savings” with dataset-dependent runtime, peak memory and per-volume time on AMOS (and ideally ACDC) under the same protocol are needed to support generalisation of the memory claim and to map the sparsity–runtime relationship beyond the three reported sets.
minor comments (6)
  1. [§3.3] §3.3 states that gradients are stopped through injected ancestor keys/values to avoid write contention, with the claim that ancestors remain trainable as queries/local neighbours. This design choice is not ablated; a brief note or small control (detach vs no-detach) would clarify whether hierarchical context learning is impaired.
  2. [Table 6] Table 6 component removals and cross-scale variants are single-fold KiTS (and AMOS only for panel a). Even standard deviations across seeds or a second fold would make the 0.7–1.0 Dice drops easier to interpret.
  3. [§2.3] Related work (§2.3) discusses ARTA, TokenSeg, HRViT, and AutoFocusFormer conceptually only. A short table contrasting fixed vs input-dependent token count, single vs mixed resolution per stage, and sequential vs parallel selection would help readers place BATS without implying unavailable numerical comparisons.
  4. [§3.3, Eq. (1)] Eq. (1): w_s = σ(ℓ_{s−1}) is described as the parent split probability; confirm whether ℓ is the pre- or post-sigmoid logit and whether w_s is detached from the discrete cascade decisions used at inference.
  5. [Abstract, §1] Minor typography: Abstract and §1 have missing spaces after periods in several places in the source (“memoryandinference”, “Manyhigh-performing”); clean for camera-ready.
  6. [Figure 3] Figure 3 token colours (darker = coarser) are helpful; adding an approximate retained-token count per example would link the qualitative panels to Table 3.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture paper scored on external public benchmarks with independent baselines.

full rationale

BATS is an engineering/architecture paper whose load-bearing claims are measured Dice, peak GPU memory, and inference time under the standardised nnU-Net Revisited protocol on five public CT/MRI datasets. Baselines and protocol numbers are taken from an external benchmark (nnU-Net Revisited / MedNeXt), not from a self-defined uniqueness theorem or a fitted constant renamed as prediction. Boundary heads are supervised with Dice-BCE against ground-truth multi-class patch occupancy (max≠min labels); the segmentation head is supervised with Dice-CE against voxel labels; ablations (parallel vs iterative vs random hierarchy, parent cluster attention, component removals) compare concrete design choices on held-out folds. Citation of ARTA is prior-method lineage for 2D adaptive tokens and is not used to force the 3D accuracy or memory results. There is no equation chain in which a claimed prediction reduces by construction to its fitted inputs, and no self-citation that substitutes for external evidence. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing content is mostly standard deep-learning and medical-segmentation practice plus a few design choices that define BATS. Free parameters are the usual architectural and training knobs (threshold, warmup, loss weights, hierarchy sizes). Axioms are domain assumptions about where resolution matters and that the nnU-Net Revisited protocol is a fair external yardstick. Invented entities are the named mechanisms, which are engineering constructs with empirical handles inside the paper’s ablations, not new physical objects.

free parameters (5)
  • split threshold τ = 0.5
    Binary keep/split decision on boundary probability; fixed at 0.5 after noting post-training bimodal scores. Directly controls token density and thus memory/runtime/accuracy trade-off.
  • oracle warmup schedule q(t) = 0.8→0 over first 25% of training
    Probability of feeding ground-truth hierarchies decays linearly 0.8→0 over first 25% of training; chosen to stabilize early refiner training, not derived.
  • auxiliary loss weights = 0.15 each
    Each of two class-fraction auxiliary heads weighted 0.15 in the total loss; hand-set relative to unit segmentation and boundary losses.
  • token scale set and widths = scales {16,8,4,2,1}; refiner widths 512…32; m=8
    Ns=5 cubic scales {16,8,4,2,1} with specified widths/blocks/heads/cluster size m=8 and neighborhood sizes; capacity allocation chosen by design, not identified from data.
  • learning rates and weight decay split = lr=1e-3; wd=1e-2 / 1e-4
    AdamW lr 1e-3, wd 1e-2 refiner / 1e-4 boundary predictor; standard but free training hyperparameters affecting the reported operating point.
axioms (5)
  • domain assumption Patches whose labels are not constant (max≠min) are the correct supervision target for where finer tokens should be allocated.
    §3.2 defines binary boundary targets from multi-class occupancy; the whole sparsity mechanism assumes class transitions proxy needed spatial detail.
  • ad hoc to paper Retaining ancestor chains (fine-first cascade) plus parent injection suffices for cross-scale context without dense multi-scale refiner feature maps.
    §3.3 design premise; supported by ablation vs local-only and separate cross-attention but not proven necessary in general.
  • domain assumption nnU-Net Revisited preprocessing, splits, training length, and sliding-window protocol provide a fair external comparison to published baselines.
    §4.1 adopts the benchmark wholesale and takes baseline scores from it; central relative claims depend on this yardstick.
  • ad hoc to paper Stopping gradients through injected ancestor key/value copies does not materially harm learning of hierarchical context.
    §3.3 engineering choice to avoid write contention; forward path unchanged, but backward path is asymmetric by construction.
  • standard math Standard Dice/CE segmentation objectives and ConvNeXt/FPN/attention building blocks behave as in prior literature.
    Uses established losses and blocks (MedNeXt-style, AutoFocusFormer clustering, RoPE3D) as black-box components.
invented entities (3)
  • Fine-first context cascade no independent evidence
    purpose: Convert parallel multi-scale split scores into a sparse hierarchy that always keeps ancestors of selected fine tokens, avoiding irreversible coarse gating misses.
    Named mechanism central to BATS vs iterative ARTA-style selection; validated by recall and Dice ablations on KiTS, not an external natural object.
  • Parent cluster attention no independent evidence
    purpose: Inject deterministic hierarchical ancestors into each token’s local cluster neighborhood for cross-scale context without dense maps or token-level NN search.
    Hybrid of cluster attention and hierarchy links; ablation shows ~1 Dice gain on KiTS vs local clusters alone.
  • BATS mixed-resolution token hierarchy (input-dependent) independent evidence
    purpose: Represent each crop with variable count and spatial mix of scales driven by boundary predictions.
    The operational representation whose sparsity is claimed to yield memory savings; statistics in Table 3 are the empirical handle.

pith-pipeline@v1.2.0-daily-grok45 · 20670 in / 4022 out tokens · 72029 ms · 2026-07-30T19:50:16.606745+00:00 · methodology

0 comments
read the original abstract

Many high-performing volumetric segmentation models maintain dense multi-scale feature maps, leading to high activation memory and inference cost. We present BATS (Boundary-Aware Token Selection), a 3D medical image segmentation architecture that concentrates fine-resolution processing near predicted class boundaries. A dense boundary predictor identifies where additional resolution is needed, while a fine-first context cascade constructs an input-dependent mixed-resolution hierarchy. Homogeneous regions are represented coarsely, with finer tokens retained around boundaries, thin structures, and small targets. The sparse hierarchy is refined and rasterised into a dense segmentation. BATS predicts boundary relevance independently at every resolution level, preventing an erroneous coarse-scale decision from suppressing fine-scale evidence. Parent cluster attention further injects hierarchical ancestor tokens into local attention neighbourhoods, providing cross-scale context without dense multi-scale feature maps or cross-scale neighbour search. We evaluate BATS on five public CT and MRI datasets using the standardised nnU-Net Revisited protocol. BATS achieves the highest LiTS Dice among the compared methods and averages within 0.37 Dice points of the strongest dense baseline, MedNeXt-L, across the five datasets. Relative to MedNeXt-L, it reduces peak allocated GPU memory by more than 53% on KiTS, LiTS, and BraTS. Inference is up to 30% faster on KiTS and LiTS, which retain fewer tokens, but slower on the more token-dense BraTS. Mixed-resolution processing therefore provides consistent memory savings, while runtime and accuracy gains depend on dataset boundary density.

Figures

Figures reproduced from arXiv: 2607.26829 by David Hagerman, Fredrik Kahl, Roman Naeem.

Figure 1
Figure 1. Figure 1: BATS architecture. A dense CNN predicts split scores at the four patch scales (163 , 8 3 , 4 3 , 2 3 ). The fine-first context cascade converts these scores into a mixed￾resolution token hierarchy and retains each selected token’s ancestors. The refiner combines FPN and raw-patch features and processes the hierarchy using parent clus￾ter attention, which augments each token’s local neighbourhood with injec… view at source ↗
Figure 2
Figure 2. Figure 2: Parent cluster attention block and mechanism. Retained ancestors augment the local neighbourhood, with detached ancestor keys contributing to the shared attention softmax and the corresponding values to the weighted aggregation. neighbourhood Nlocal(i). Nearest-neighbour search is therefore performed only over cluster centroids rather than over all tokens, while overlapping neighbour￾hoods permit informati… view at source ↗
Figure 3
Figure 3. Figure 3: Representative predictions and retained token hierarchies on ACDC, LiTS, BraTS, KiTS, and AMOS. From left to right: ground truth, prediction, token selection, zoomed prediction, and zoomed selection. Retained tokens are shown as filled teal patches, with darker colours indicating coarser tokens. Fine-resolution tokens generally follow class boundaries [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Failure cases on LiTS, KiTS, and AMOS. From left to right: ground truth, prediction, GT-based token selection, predicted token selection, and a zoomed crop showing the ground-truth segmentation over the predicted selection. Token colours follow [PITH_FULL_IMAGE:figures/full_fig_p020_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

35 extracted references · 7 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2107.02314 (2021)

    Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classifica- tion. arXiv preprint arXiv:2107.02314 (2021)

    Pith/arXiv arXiv 2021

  2. [2]

    Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging37(11), 2514–2525 (2018)

    2018

  3. [3]

    Medical image analysis84, 102680 (2023)

    Bilic,P.,Christ,P.,Li,H.B.,Vorontsov,E.,Ben-Cohen,A.,Kaissis,G.,Szeskin,A., Jacobs, C., Mamani, G.E.H., Chartrand, G., et al.: The liver tumor segmentation benchmark (lits). Medical image analysis84, 102680 (2023)

    2023

  4. [4]

    arXiv preprint arXiv:2210.09461 (2022) BATS: Boundary-Aware Mixed-Resolution Tokens 21

    Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022) BATS: Boundary-Aware Mixed-Resolution Tokens 21

    Pith/arXiv arXiv 2022

  5. [5]

    In: International Conference on Medical Im- age Computing and Computer-Assisted Intervention

    Guo, Y., Song, B., Fan, H., Cheng, E.: Edge-aware token halting for efficient and accurate medical image segmentation. In: International Conference on Medical Im- age Computing and Computer-Assisted Intervention. pp. 185–195. Springer (2025)

    2025

  6. [6]

    arXiv preprint arXiv:2603.26258 (2026)

    Hagerman, D., Naeem, R., Brorsson, E., Kahl, F., Svensson, L.: Arta: Adap- tive mixed-resolution token allocation for efficient dense feature extraction. arXiv preprint arXiv:2603.26258 (2026)

    arXiv 2026

  7. [7]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    He, Y., Nath, V., Yang, D., Tang, Y., Myronenko, A., Xu, D.: Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 416–426. Springer (2023)

    2023

  8. [8]

    In: CVPR

    He, Y., Yang, D., Roth, H., Zhao, C., Xu, D.: Dints: Differentiable neural network topology search for 3d medical image segmentation. In: CVPR. pp. 5841–5850 (June 2021)

    2021

  9. [9]

    arXiv preprint arXiv:2307.01984 (2023)

    Heller, N., Isensee, F., Trofimova, D., Tejpaul, R., Zhao, Z., Chen, H., Wang, L., Golts, A., Khapun, D., Shats, D., et al.: The kits21 challenge: Automatic segmen- tation of kidneys, renal tumors, and renal cysts in corticomedullary-phase ct. arXiv preprint arXiv:2307.01984 (2023)

    Pith/arXiv arXiv 2023

  10. [10]

    arXiv preprint arXiv:2304.06716 (2023)

    Huang, Z., Wang, H., Deng, Z., Ye, J., Su, Y., Sun, H., He, J., Gu, Y., Gu, L., Zhang, S., et al.: Stu-net: Scalable and transferable medical image segmen- tation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716 (2023)

    Pith/arXiv arXiv 2023

  11. [11]

    Nature methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

    2021

  12. [12]

    In: International Conference on Medical Image Computing and Computer- Assisted Intervention

    Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 488–498. Springer (2024)

    2024

  13. [13]

    Advances in neural information processing systems 35, 36722–36732 (2022)

    Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhanng, L., Ma, W., Wan, X., 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

  14. [14]

    In: Proceed- ings of the MICCAI Multi-Atlas Labeling Beyond the Cranial Vault Workshop and Challenge

    Landman, B., Xu, Z., Iglesias, J.E., Styner, M., Langerak, T., Klein, A.: MICCAI multi-atlas labeling beyond the cranial vault: Workshop and challenge. In: Proceed- ings of the MICCAI Multi-Atlas Labeling Beyond the Cranial Vault Workshop and Challenge. vol. 5, p. 12. Munich, Germany (2015)

    2015

  15. [15]

    IEEE Journal of Biomedical and Health Informatics (2023)

    Lin, X., Yu, L., Cheng, K.T., Yan, Z.: Batformer: Towards boundary-aware lightweight transformer for efficient medical image segmentation. IEEE Journal of Biomedical and Health Informatics (2023)

    2023

  16. [16]

    IEEE transactions on medical imaging42(8), 2325–2337 (2023)

    Lin, X., Yu, L., Cheng, K.T., Yan, Z.: The lighter the better: rethinking transform- ers in medical image segmentation through adaptive pruning. IEEE transactions on medical imaging42(8), 2325–2337 (2023)

    2023

  17. [17]

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

    2021

  18. [18]

    arXiv preprint arXiv:2401.04722 (2024)

    Ma, J., Li, F., Wang, B.: U-mamba: Enhancing long-range dependency for biomed- ical image segmentation. arXiv preprint arXiv:2401.04722 (2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Hagerman et al

    Menze, B., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor 22 D. Hagerman et al. image segmentation benchmark (brats). IEEE transactions on medical imaging 34(10), 1993–2024 (2014)

    1993

  20. [20]

    In: International MICCAI brainlesion workshop

    Myronenko,A.:3dmribraintumorsegmentationusingautoencoderregularization. In: International MICCAI brainlesion workshop. pp. 311–320. Springer (2018)

    2018

  21. [21]

    In: Kidney and Kidney Tumor Segmentation

    Myronenko, A., Yang, D., He, Y., Xu, D.: Automated 3d segmentation of kid- neys and tumors in MICCAI KiTS 2023 challenge. In: Kidney and Kidney Tumor Segmentation. Lecture Notes in Computer Science, vol. 14540, pp. 1–7. Springer, Cham (2024)

    2023

  22. [22]

    arXiv preprint arXiv:1804.03999 (2018)

    Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)

    Pith/arXiv arXiv 2018

  23. [23]

    Advances in neural infor- mation processing systems34, 13937–13949 (2021)

    Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., Hsieh, C.J.: Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural infor- mation processing systems34, 13937–13949 (2021)

    2021

  24. [24]

    In: International conference on medical image computing and computer-assisted intervention

    Roy, S., Koehler, G., Ulrich, C., Baumgartner, M., Petersen, J., Isensee, F., Jaeger, P.F., Maier-Hein, K.H.: Mednext: transformer-driven scaling of convnets for med- ical image segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 405–415. Springer (2023)

    2023

  25. [25]

    Medical image analysis53, 197–207 (2019)

    Schlemper, J., Oktay, O., Schaap, M., Heinrich, M., Kainz, B., Glocker, B., Rueck- ert, D.: Attention gated networks: Learning to leverage salient regions in medical images. Medical image analysis53, 197–207 (2019)

    2019

  26. [26]

    IEEE Transac- tions on Medical Imaging43(9), 3377–3390 (2024)

    Shaker, A., Maaz, M., Rasheed, H., Khan, S., Yang, M.H., Khan, F.S.: Unetr++: delving into efficient and accurate 3d medical image segmentation. IEEE Transac- tions on Medical Imaging43(9), 3377–3390 (2024)

    2024

  27. [27]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swin transformers for 3d med- ical image analysis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20730–20740 (2022)

    2022

  28. [28]

    IEEE Transactions on Medical Imaging32(2), 178–188 (2013)

    Ulén, J., Strandmark, P., Kahl, F.: An efficient optimization framework for multi- region segmentation based on lagrangian duality. IEEE Transactions on Medical Imaging32(2), 178–188 (2013)

    2013

  29. [29]

    In: International conference on medical image computing and computer-assisted intervention

    Xie, Y., Zhang, J., Shen, C., Xia, Y.: Cotr: Efficiently bridging cnn and transformer for 3d medical image segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 171–180. Springer (2021)

    2021

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10809–10818 (2022)

    2022

  31. [31]

    arXiv preprint arXiv:2601.04519 (2026)

    Zeng, S., Zhou, H., Zhu, Z., Liu, Y.: Tokenseg: Efficient 3d medical image segmen- tation via hierarchical visual token compression. arXiv preprint arXiv:2601.04519 (2026)

    arXiv 2026

  32. [32]

    IEEE transactions on image processing32, 4036–4045 (2023)

    Zhou, H.Y., Guo, J., Zhang, Y., Han, X., Yu, L., Wang, L., Yu, Y.: nnformer: Volumetric medical image segmentation via a 3d transformer. IEEE transactions on image processing32, 4036–4045 (2023)

    2023

  33. [33]

    In: International Conference on Information Processing in Medical Imaging

    Zhou, L., Liu, H., Bae, J., He, J., Samaras, D., Prasanna, P.: Token sparsification for faster medical image segmentation. In: International Conference on Information Processing in Medical Imaging. pp. 743–754. Springer (2023)

    2023

  34. [34]

    arXiv preprint arXiv:2010.04159 (2020) BATS: Boundary-Aware Mixed-Resolution Tokens 23

    Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) BATS: Boundary-Aware Mixed-Resolution Tokens 23

    Pith/arXiv arXiv 2010

  35. [35]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ziwen, C., Patnaik, K., Zhai, S., Wan, A., Ren, Z., Schwing, A.G., Colburn, A., Fuxin, L.: Autofocusformer: Image segmentation off the grid. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18227– 18236 (2023)

    2023