Axial-Centric Cross-Plane Attention for 3D Medical Image Classification

Doyoung Park; Jinsoo Kim; Lohendran Baskaran

REVIEW 3 major objections 5 minor 26 references

Three planes, one anchor: axial-centric cross-plane attention for 3D medical image classification.

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 · deepseek-v4-flash

2026-08-02 20:57 UTC pith:PAYM4IRK

load-bearing objection Useful architectural idea with a clean ablation story, but the headline claim outruns the evidence because every comparison is against baselines that lack the MedDINOv3 backbone. the 3 major comments →

arxiv 2602.21636 v2 pith:PAYM4IRK submitted 2026-02-25 cs.CV

Axial-Centric Cross-Plane Attention for 3D Medical Image Classification

Doyoung Park , Jinsoo Kim , Lohendran Baskaran This is my paper

classification cs.CV

keywords axial-centric attentioncross-plane attention3D medical image classificationmulti-planar representation learningMedMNIST3Dvision foundation modelscomputed tomographytransformer

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.

This paper proposes that 3D medical image classifiers should be built around the same asymmetry radiologists actually use: the axial plane is the primary reference, and the coronal and sagittal planes supply supporting context. The authors implement this as a pipeline where a frozen medical vision model extracts 2D slice features from each plane, per-plane transformers build global embeddings, and two cross-plane attention modules let axial features query coronal and sagittal features as key and value. Tested on six MedMNIST3D datasets, the paper reports the best accuracy on five of six and the highest AUC on three of six, with a lightweight variant staying competitive. The paper interprets these results as evidence that aligning architecture with clinical reading patterns helps more than adding parameters.

Core claim

The paper's central claim is that asymmetric inter-plane attention—axial as query, coronal and sagittal as key/value—yields better 3D medical image classification than equal-weight multi-plane fusion or full-volume 3D processing. The architecture keeps the axial feature stream as the reference, refines it through two separate cross-attention encoders conditioned on coronal and sagittal embeddings, omits the residual connection after cross-attention because the two sides are semantically distinct, and averages logits from two MLP heads. On the six MedMNIST3D datasets, the proposal reports the highest ACC in five cases and highest AUC in three, with its closest margins on Fracture and Adrenal.

What carries the argument

The load-bearing component is the axial-centric cross-plane transformer encoder. Each plane's slice features come from a frozen MedDINOv3 backbone, then RICA blocks (positional attention modules) and a 12-layer transformer encoder produce a global embedding per plane. In the cross-plane encoder, the axial embedding is used only as the query, the coronal or sagittal embedding only as keys and values, and the residual connection after cross-attention is dropped. That directional query-key-value arrangement, plus two independent axial-to-auxiliary fusions rather than one sequential chain, is what operationalizes the clinical asymmetry.

Load-bearing premise

The clean-comparison claim rests on the numbers in Table 1 being fair comparators, but those rows come from separate publications trained under their own pipelines, so if the frozen MedDINOv3 backbone or per-dataset hyperparameters account for the accuracy gap, the paper's attribution to axial-centric cross-plane attention is unsupported.

What would settle it

Retrain the six compared baselines under the paper's exact protocol—same frozen MedDINOv3 backbone, same per-dataset learning rates and augmentations, same epochs—and observe whether the axial-centric model still wins. If any baseline then matches or beats it, the reported gain cannot be attributed to axial-centric cross-plane attention.

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

If this is right

If axial-centric querying is the cause of the gains, then symmetric multi-plane fusion is leaving measurable accuracy on the table, and re-centering any multi-planar model on a fixed primary plane should improve it at no extra parameter cost.
The competitive lightweight variant implies a frozen medical foundation extractor plus a modest trainable head can approach full-size 3D models, which matters when compute or labeled data is scarce.
The ablation results give concrete design rules for other multi-plane architectures: keep the primary plane as query, fuse auxiliary planes in parallel rather than sequentially, and drop the cross-attention residual.
Because the method works across CT, MRA, and electron microscopy volumes in the benchmark, the axial-centric principle is not tied to one imaging modality.

Where Pith is reading between the lines

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

The paper credits the architecture, but its comparison does not retrain the six baselines with the same frozen MedDINOv3 backbone and per-dataset schedules; an equally plausible reading is that the pretrained features drive most of the gain. Retraining baselines under the paper's protocol would separate the two.
The axial-centric idea is probably general: any modality with a conventionally primary plane (e.g., long-axis in echocardiography or the sagittal plane in some MRI protocols) should see the same benefit when that plane is made the query. The paper only tests volumes where axial is standard.
A decisive extension would be to fix the architecture and permute which plane acts as query; the authors' reasoning predicts the fixed primary plane should win regardless of which plane it is, and equalizing the two directions should fall in between.
The two auxiliary streams never interact directly—context is mixed only through the axial representation. Whether allowing coronal and sagittal streams to exchange information before fusion helps is an open, testable question the paper does not address.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Useful architectural idea with a clean ablation story, but the headline claim outruns the evidence because every comparison is against baselines that lack the MedDINOv3 backbone.

read the letter

Short version: this is a real architectural proposal with a clean ablation story, but the headline claim is overstated and the main attribution is not yet supported. The combination of frozen MedDINOv3 features per plane, RICA blocks, independent intra-plane transformers, and asymmetric axial-centric cross-attention with fixed axial queries is new, and the ablation that reverses QKV allocation is a useful contribution. The paper is also honest enough to include ablations for capacity and fusion order, and the benchmark results are plausible.

The problems are proportionate to the claims. First, the abstract's 'consistently outperforms... in ACC and AUC' is not what Table 1 shows: Ours is second-best on NoduleMNIST3D ACC and AUC, and second-best on VesselMNIST3D AUC. That is a factual overstatement, easy to fix. Second, and more substantively, the stress-test note is right: there is no same-backbone control. All baselines are published numbers trained without MedDINOv3. Since the ablations always keep the frozen backbone, they never test whether the cross-plane attention itself contributes beyond the VFM features. A reader cannot tell if the gains come from the pretrained backbone or the axial-centric design. Adding a simple head on MedDINOv3 features, or retraining one or two baselines under the same protocol, would answer this. Third, AC-Tiny is mentioned in the abstract but never defined in the main text, so that claim is unverifiable. Fourth, no error bars or multiple seeds, so we do not know if the gaps are noise.

None of this is fatal to the core idea. The architecture is clearly specified, the ablation isolating QKV allocation is informative, and the external VFM choice is reasonable. The circularity burden is low. But the central claim needs stronger evidence.

Who should read it: anyone working on multi-planar medical image classification, especially with VFMs. It deserves a serious referee, though I would expect major revision. The referee should ask for a same-backbone control and clarification of AC-Tiny. I would recommend sending it to peer review rather than desk rejection.

Referee Report

3 major / 5 minor

Summary. The paper proposes an axial-centric cross-plane attention architecture for 3D medical image classification. Volumes are resliced into axial, coronal, and sagittal planes; each plane is encoded by the frozen MedDINOv3 vision foundation model, then processed by RICA positional blocks and intra-plane transformer encoders. The axial representation is then conditioned on coronal and sagittal representations through two axial-centric cross-plane transformer encoders, and the CLS tokens from the two fused streams are passed to MLP heads whose logits are averaged. The method is evaluated on all six MedMNIST3D datasets, reporting the best accuracy on five and the best AUC on three datasets, with ablations supporting the axial-centric QKV allocation and dual independent cross-plane fusion. The abstract additionally claims a lightweight AC-Tiny variant, though no such model is defined or reported in the main text.

Significance. If the empirical claims are supported, the paper would make a useful contribution by showing that an axial-first clinical prior can be encoded in a cross-plane attention mechanism and that frozen medical VFM features can support data-efficient 3D classification. The evaluation covers the full MedMNIST3D benchmark, and the ablation study isolates several design choices. The work is not circular: the RICA block is a self-cited prior component but is part of the architecture rather than the defining quantity of the headline claim, and MedDINOv3 is an external pretrained model. The main value of the paper is conditional on resolving the comparison protocol, because the reported gains cannot currently be attributed to the proposed attention design rather than to the frozen backbone or per-dataset hyperparameter choices.

major comments (3)

[Sec. 3.2 and Table 1; Sec. 2.2] The central comparison is confounded. 'Ours' uses a frozen MedDINOv3 pretrained on 3.87M axial CT images, while every baseline row in Table 1 is taken from separate publications trained under their own pipelines; none uses this backbone. The reported gaps (e.g., +4.4 ACC on AdrenalMNIST3D, +3.9 ACC on FractureMNIST3D) could therefore be due to the pretrained feature extractor, or to per-dataset learning rates/augmentation choices described in Sec. 3.1, rather than to axial-centric cross-plane attention. The ablations in Table 2 keep the same frozen backbone throughout and vary only capacity, fusion order, and QKV allocation, so they never measure the contribution of the cross-plane attention relative to the backbone alone. I request a same-backbone control — MedDINOv3 features with a simple head, and preferably an axial-only transformer without cross-plane fusion — trained under the iden
[Abstract; Table 1] The abstract's claim that the architecture 'consistently outperforms existing 3D and multi-plane models in ACC and AUC' is contradicted by the paper's own Table 1: on NoduleMNIST3D the method is second-best in ACC (89.0 vs 90.3) and second-best in AUC (90.2 vs 94.3), and on VesselMNIST3D it is second-best in AUC (93.3 vs 95.9). The claim should be revised to match the actual win/loss pattern or be backed by statistical testing. In addition, the abstract credits 'AC-Tiny' with competitive lightweight performance, but this variant is never defined, sized, or tabulated anywhere in the manuscript; this claim is unverifiable and should be either implemented/reported or removed.
[Sec. 3.1; Tables 1 and 2] No uncertainty quantification is provided. All six MedMNIST3D datasets are small (1,370–1,908 volumes), and all results are point estimates from a single run. Without multiple seeds (mean±std) or significance testing, the performance differences in Table 1 — some only 0.2–1.3 points — and the small ablation differences in Table 2 (e.g., 60.8 vs 60.4 ACC on FractureMNIST3D; 86.1 vs 83.5–84.8 on SynapseMNIST3D) could be within run-to-run noise. I ask for at least 3–5 seeds and, for the central comparisons, paired tests or effect-size reporting.

minor comments (5)

[Sec. 3.2, Table 1] The baseline name 'CdTrandformer + CcCL' appears to be a typo for 'CdTransformer + CcCL'; the same typo occurs in Ref. [24].
[Sec. 2.2 / Fig. 1] The figure legend is dense: 'RICA Block', 'Transformer Encoder', and 'Cross-plane Transformer Encoder' are used with multiple meanings. Consider labeling the three plane-specific branches and the two cross-plane branches with distinct numerals to make the data flow clearer.
[Sec. 2.2] MedDINOv3 is pretrained on axial CT slices, but it is applied here to coronal and sagittal reslices. The resulting distribution shift is not discussed; a brief justification or an ablation on this choice would be helpful.
[Sec. 3.1] The text mentions 'MedMNIST+ benchmark' but the reference list only includes MedMNIST v2 [16] and the ISBI 2021 paper [26]. Please clarify the relationship or add the appropriate citation.
[Sec. 3.3] The ablation in Table 2 does not include a variant without RICA blocks or a variant without the cross-plane encoders (e.g., axial-only). Such variants would strengthen the attribution of the design choices, though they are secondary to the same-backbone control requested above.

Circularity Check

0 steps flagged

No circularity found; the architecture is empirically evaluated, and comparison-fairness issues are not circularity.

full rationale

The paper's central claim is an empirical architecture comparison, not an analytic derivation whose outputs reduce to its inputs by construction. Table 1 compares the proposed model to published baselines from other pipelines; the absence of a same-backbone control is a validity/attribution concern (does the frozen MedDINOv3 backbone or per-dataset tuning explain the gains?), but it is not circular because the reported accuracies and AUCs are measured, not constructed from the model definition. Table 2 ablations vary QKV allocation, fusion order, and capacity within the same pipeline; these are controlled experiments, not fitted parameters renamed as predictions. The RICA block is cited from the first author's prior work, but it is a modular component whose use is supported by a published external paper and it is not used to justify the central axial-centric cross-plane claim via an unverified uniqueness theorem; this self-citation is not load-bearing in a circular sense. The abstract's AC-Tiny lightweight claim is not reported in the main text, which is a missing-support issue rather than a circular reduction. No step in the paper equates a definition with a result or imports a conclusion solely through self-citation, so no significant circularity is present.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

No new physical entities or fitted physical constants are introduced. The free parameters are training and architecture choices rather than parameters of a scientific law. The RICA block is self-cited but independently ablatable, so it is not a hidden dependency that forces the result.

free parameters (3)

Per-dataset maximum learning rate = 1e-4 (NoduleMNIST3D); 1e-5 (all others)
Section 3.1 sets the max LR per dataset; every reported result is produced under one of these two hand-chosen values, introducing per-dataset tuning.
Selective data augmentation schedule = Unspecified subset of RandomAnisotropy, RandomAffine, RandomFlip, RandomNoise, RandomBlur, RandomGamma
Section 3.1 says six TorchIO methods were 'selectively applied', but does not state which datasets received which transforms; this is a hand-tuned protocol affecting all results.
Transformer depth and heads = N=12, H=12 full model; N=6, H=4 reduced model
Architecture capacity is chosen by hand; the capacity-ablation claim in Table 2 depends on these specific values.

axioms (3)

domain assumption MedDINOv3 features trained on axial CT slices transfer to coronal/sagittal reformations and to non-CT modalities such as MRA and electron microscopy.
Section 2.2 applies the same frozen extractor to every plane of every dataset, including VesselMNIST3D (MRA) and SynapseMNIST3D (EM), with no domain adaptation or analysis.
domain assumption Previously published baseline metrics are direct comparators without retraining under the authors' protocol.
Table 1 mixes numbers from refs [16,20-24] with the authors' own pipeline; no common training budget, pretraining, or augmentation is shared across rows.
domain assumption A 64-slice sequence of 2D embeddings per plane preserves enough 3D spatial context for classification.
Sections 2.2-2.3 collapse each plane to B x D x E tokens with 1D positional embeddings; no 3D inductive bias or depth-aware encoding is evaluated.

pith-pipeline@v1.3.0-alltime-deepseek · 7167 in / 14203 out tokens · 133902 ms · 2026-08-02T20:57:39.504054+00:00 · methodology

0 comments

read the original abstract

Abridged: Clinicians commonly interpret 3D medical images by examining multiple anatomical planes rather than relying on volumetric views. In clinical CT workflows, the axial plane often serves as the primary diagnostic reference, while the auxiliary planes provide complementary spatial context. However, many existing 3D deep learning approaches either process volumetric data holistically or assign equal importance to all planes, failing to reflect this asymmetric, axial-centric interpretation strategy. To address this, we propose an axial-centric cross-plane attention architecture for 3D medical image classification that models asymmetric dependencies between anatomical planes. The architecture employs large-scale axial CT images pretrained MedDINOv3 as a frozen feature extractor for axial, coronal, and sagittal planes. RICA blocks and intra-plane transformer encoders capture plane-specific positional and contextual information, while axial-centric cross-plane transformer encoders selectively condition axial representations on complementary auxiliary representations. Experiments on six datasets from the MedMNIST3D benchmark show that the proposed method consistently outperforms existing 3D and multi-plane models in ACC and AUC. A lightweight variant, AC-Tiny, achieves competitive performance with substantially fewer trainable parameters, suggesting that architectural design contributes more to performance gains than increased model scale. Ablation studies further validate the importance of axial-centric querying, QKV allocation, directional cross-plane fusion, residual-free cross-attention, and classification head design. Slice-level Grad-CAM visualizations demonstrate that the model identifies diagnostically relevant regions across all planes. These findings highlight the value of aligning architectural design with clinical interpretation workflows for robust 3D medical image analysis.

Figures

Figures reproduced from arXiv: 2602.21636 by Doyoung Park, Jinsoo Kim, Lohendran Baskaran.

**Figure 1.** Figure 1: Overview of the proposed axial-centric cross-plane attention architecture. The architecture comprises a MedDINOv3 backbone for plane-wise feature extraction, RICA blocks for positional context modeling within each stacked feature sequence, transformer encoders for intra-plane contextual self-attention, and axial-centric crossplane transformer encoders for conditioning axial features on complementary coron… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 2 linked inside Pith

[1]

Ter-Pogossian, M.M.: Basic principles of computed axial tomography. Semin. Nucl. Med.7(2), 109–127 (1977)

1977
[2]

Mang, T., et al.: Comparison of axial, coronal, and primary 3D review in MDCT colonography for the detection of small polyps: a phantom study. Eur. J. Radiol. 70(1), 86–93 (2009)

2009
[3]

IEEE Trans

Setio, A.A.A., et al.: Pulmonary nodule detection in CT images: false positive re- duction using multi-view convolutional networks. IEEE Trans. Med. Imaging35(5), 1160–1169 (2016) Axial-Centric Cross-Plane Attention for 3D Medical Image Classification 9

2016
[4]

In: de Bruijne, M., et al

Van Tulder, G., Tong, Y., Marchiori, E.: Multi-view analysis of unregistered medical images using cross-view transformers. In: de Bruijne, M., et al. (eds.) MICCAI 2021, LNCS, vol. 12903, pp. 104–113. Springer, Cham (2021)

2021
[5]

In: Proc

Black, S., Souvenir, R.: Multi-view classification using hybrid fusion and mutual distillation. In: Proc. IEEE/CVF WACV, pp. 270–280 (2024)

2024
[6]

In: MICCAI 2025, LNCS, vol

Zheng, X., et al.: Xfmamba: Cross-fusion mamba for multi-view medical image clas- sification. In: MICCAI 2025, LNCS, vol. 15960, pp. 672–682. Springer, Cham (2025)

2025
[7]

McMenamin, D., Pearce, A., Klassen, M.: Visual search in abdominopelvic CT in- terpretation: accuracy and time efficiency between coronal MPR and axial images. Acad. Radiol.2(2), 164–168 (2015)

2015
[8]

Abdelhamid, H.M., et al.: Multiplane reconstruction modifies the diagnostic per- formance of CT angiography in carotid webs. Clin. Neurol. Neurosurg. 244, 108441 (2024)

2024
[9]

Sandrasegaran, K., et al.: Benefits of routine use of coronal and sagittal reformations in multi-slice CT examination of the abdomen and pelvis. Clin. Radiol.62(4), 340– 347 (2007)

2007
[10]

In: Proc

Jang, J., Hwang, D.: M3t: three-dimensional medical image classifier using multi- plane and multi-slice transformer. In: Proc. IEEE/CVF CVPR, pp. 20718–20729 (2022)

2022
[11]

Angkoso, C.V., et al.: Multi-features fusion in multi-plane MRI images for Alzheimer’s disease classification. Int. J. Intell. Eng. Syst.15(4), 182–197 (2022)

2022
[12]

SPIE 12927, Medical Imaging 2024: Computer-Aided Diagnosis, 1292735 (2024)

Li,Y.,Hu,M.,Yang,X.:Polyp-sam:Transfersamforpolypsegmentation.In:Proc. SPIE 12927, Medical Imaging 2024: Computer-Aided Diagnosis, 1292735 (2024)

2024
[13]

arXiv preprint arXiv:2509.06467 (2025)

Liu, C., et al.: Does DINOv3 Set a New Medical Vision Standard? Benchmark- ing 2D and 3D Classification, Segmentation, and Registration. arXiv preprint arXiv:2509.06467 (2025)

arXiv 2025
[14]

arXiv preprint arXiv:2312.02366 (2023)

Baharoon, M., et al.: Evaluating general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks. arXiv preprint arXiv:2312.02366 (2023)

Pith/arXiv arXiv 2023
[15]

arXiv preprint arXiv:2509.02379 (2025)

Li, Y., et al.: MedDINOv3: How to adapt vision foundation models for medical image segmentation?. arXiv preprint arXiv:2509.02379 (2025)

arXiv 2025
[16]

Yang, J., et al.: MedMNIST v2-a large-scale lightweight benchmark for 2D and 3D biomedical image classification. Sci. Data10(1), 41 (2023)

2023
[17]

In: 2025 IEEE 22nd ISBI, pp

Park, D., et al.: RICAU-Net: Residual-block Inspired coordinate attention U-Net for segmentation of small and sparse calcium lesions in cardiac CT. In: 2025 IEEE 22nd ISBI, pp. 1–5 (2025)

2025
[18]

Park, D., et al.: Evaluating the generalizability of an automated coronary artery calcium segmentation and scoring algorithm using multi-vendor dataset. Sci. Rep. 15(1), 21744 (2025)

2025
[19]

arXiv preprint arXiv:2010.11929 (2020)

Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. arXiv preprint arXiv:2010.11929 (2020)

Pith/arXiv arXiv 2010
[20]

Sensors24(23), 7511 (2024)

Zhu, Y., et al.: BSDA: bayesian random semantic data augmentation for medical image classification. Sensors24(23), 7511 (2024)

2024
[21]

In: Proc

Lai, Z., et al.: Residual-based language models are free boosters for biomedical imaging tasks. In: Proc. IEEE/CVF CVPR, pp. 5086–5096 (2024)

2024
[22]

Wei, H., et al.: LMTTM-VMI: Linked Memory Token Turing Machine for 3D volumetric medical image classification. Comput. Methods Programs Biomed. 262, 108640 (2025)

2025
[23]

Procedia Comput

Simionescu, C.: Medformer: A Multitask Multimodal Foundational Model for Med- ical Imaging. Procedia Comput. Sci. 270, 446–455 (2025) 10 D. Park et al

2025
[24]

In: Proc

Zhu, Q., Fu, C., Li, S.: Class-consistent contrastive learning driven cross- dimensional transformer for 3D medical image classification. In: Proc. 33rd IJCAI, pp. 1807–1815 (2024)

2024
[25]

Mei, X., et al.: RadImageNet: an open radiologic deep learning research dataset for effective transfer learning. Radiol. Artif. Intell.4(5), e210315 (2022)

2022
[26]

In: 2021 IEEE 18th ISBI, pp

Yang, J., Shi, R., Ni, B.: Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In: 2021 IEEE 18th ISBI, pp. 191–195. IEEE (2021)

2021

[1] [1]

Ter-Pogossian, M.M.: Basic principles of computed axial tomography. Semin. Nucl. Med.7(2), 109–127 (1977)

1977

[2] [2]

Mang, T., et al.: Comparison of axial, coronal, and primary 3D review in MDCT colonography for the detection of small polyps: a phantom study. Eur. J. Radiol. 70(1), 86–93 (2009)

2009

[3] [3]

IEEE Trans

Setio, A.A.A., et al.: Pulmonary nodule detection in CT images: false positive re- duction using multi-view convolutional networks. IEEE Trans. Med. Imaging35(5), 1160–1169 (2016) Axial-Centric Cross-Plane Attention for 3D Medical Image Classification 9

2016

[4] [4]

In: de Bruijne, M., et al

Van Tulder, G., Tong, Y., Marchiori, E.: Multi-view analysis of unregistered medical images using cross-view transformers. In: de Bruijne, M., et al. (eds.) MICCAI 2021, LNCS, vol. 12903, pp. 104–113. Springer, Cham (2021)

2021

[5] [5]

In: Proc

Black, S., Souvenir, R.: Multi-view classification using hybrid fusion and mutual distillation. In: Proc. IEEE/CVF WACV, pp. 270–280 (2024)

2024

[6] [6]

In: MICCAI 2025, LNCS, vol

Zheng, X., et al.: Xfmamba: Cross-fusion mamba for multi-view medical image clas- sification. In: MICCAI 2025, LNCS, vol. 15960, pp. 672–682. Springer, Cham (2025)

2025

[7] [7]

McMenamin, D., Pearce, A., Klassen, M.: Visual search in abdominopelvic CT in- terpretation: accuracy and time efficiency between coronal MPR and axial images. Acad. Radiol.2(2), 164–168 (2015)

2015

[8] [8]

Abdelhamid, H.M., et al.: Multiplane reconstruction modifies the diagnostic per- formance of CT angiography in carotid webs. Clin. Neurol. Neurosurg. 244, 108441 (2024)

2024

[9] [9]

Sandrasegaran, K., et al.: Benefits of routine use of coronal and sagittal reformations in multi-slice CT examination of the abdomen and pelvis. Clin. Radiol.62(4), 340– 347 (2007)

2007

[10] [10]

In: Proc

Jang, J., Hwang, D.: M3t: three-dimensional medical image classifier using multi- plane and multi-slice transformer. In: Proc. IEEE/CVF CVPR, pp. 20718–20729 (2022)

2022

[11] [11]

Angkoso, C.V., et al.: Multi-features fusion in multi-plane MRI images for Alzheimer’s disease classification. Int. J. Intell. Eng. Syst.15(4), 182–197 (2022)

2022

[12] [12]

SPIE 12927, Medical Imaging 2024: Computer-Aided Diagnosis, 1292735 (2024)

Li,Y.,Hu,M.,Yang,X.:Polyp-sam:Transfersamforpolypsegmentation.In:Proc. SPIE 12927, Medical Imaging 2024: Computer-Aided Diagnosis, 1292735 (2024)

2024

[13] [13]

arXiv preprint arXiv:2509.06467 (2025)

Liu, C., et al.: Does DINOv3 Set a New Medical Vision Standard? Benchmark- ing 2D and 3D Classification, Segmentation, and Registration. arXiv preprint arXiv:2509.06467 (2025)

arXiv 2025

[14] [14]

arXiv preprint arXiv:2312.02366 (2023)

Baharoon, M., et al.: Evaluating general purpose vision foundation models for medical image analysis: An experimental study of dinov2 on radiology benchmarks. arXiv preprint arXiv:2312.02366 (2023)

Pith/arXiv arXiv 2023

[15] [15]

arXiv preprint arXiv:2509.02379 (2025)

Li, Y., et al.: MedDINOv3: How to adapt vision foundation models for medical image segmentation?. arXiv preprint arXiv:2509.02379 (2025)

arXiv 2025

[16] [16]

Yang, J., et al.: MedMNIST v2-a large-scale lightweight benchmark for 2D and 3D biomedical image classification. Sci. Data10(1), 41 (2023)

2023

[17] [17]

In: 2025 IEEE 22nd ISBI, pp

Park, D., et al.: RICAU-Net: Residual-block Inspired coordinate attention U-Net for segmentation of small and sparse calcium lesions in cardiac CT. In: 2025 IEEE 22nd ISBI, pp. 1–5 (2025)

2025

[18] [18]

Park, D., et al.: Evaluating the generalizability of an automated coronary artery calcium segmentation and scoring algorithm using multi-vendor dataset. Sci. Rep. 15(1), 21744 (2025)

2025

[19] [19]

arXiv preprint arXiv:2010.11929 (2020)

Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. arXiv preprint arXiv:2010.11929 (2020)

Pith/arXiv arXiv 2010

[20] [20]

Sensors24(23), 7511 (2024)

Zhu, Y., et al.: BSDA: bayesian random semantic data augmentation for medical image classification. Sensors24(23), 7511 (2024)

2024

[21] [21]

In: Proc

Lai, Z., et al.: Residual-based language models are free boosters for biomedical imaging tasks. In: Proc. IEEE/CVF CVPR, pp. 5086–5096 (2024)

2024

[22] [22]

Wei, H., et al.: LMTTM-VMI: Linked Memory Token Turing Machine for 3D volumetric medical image classification. Comput. Methods Programs Biomed. 262, 108640 (2025)

2025

[23] [23]

Procedia Comput

Simionescu, C.: Medformer: A Multitask Multimodal Foundational Model for Med- ical Imaging. Procedia Comput. Sci. 270, 446–455 (2025) 10 D. Park et al

2025

[24] [24]

In: Proc

Zhu, Q., Fu, C., Li, S.: Class-consistent contrastive learning driven cross- dimensional transformer for 3D medical image classification. In: Proc. 33rd IJCAI, pp. 1807–1815 (2024)

2024

[25] [25]

Mei, X., et al.: RadImageNet: an open radiologic deep learning research dataset for effective transfer learning. Radiol. Artif. Intell.4(5), e210315 (2022)

2022

[26] [26]

In: 2021 IEEE 18th ISBI, pp

Yang, J., Shi, R., Ni, B.: Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In: 2021 IEEE 18th ISBI, pp. 191–195. IEEE (2021)

2021