Pith. sign in

REVIEW 3 major objections 8 minor 20 references

AnomExpert is a prototype-driven framework that learns to identify anatomical planes and select disease-relevant views from case-level labels alone, reaching 86.9% accuracy and 84.2% F1 on a multi-center dataset of 3,654 prenatal ultrasound

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 ↗

Prototype-based plane grouping plus disease-aware sparse plane selection improves weakly supervised prenatal ultrasound anomaly classification by 1.4 accuracy points over the best MIL baseline.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid MIL paper for prenatal ultrasound: the empirical core is credible, but the 'anatomical plane' interpretation is not proven by the evidence. the 3 major comments →

arxiv 2607.13409 v1 pith:BD33HFHT submitted 2026-07-15 cs.CV

AnomExpert: Identifying and Selecting Anatomical Planes for Prenatal Ultrasound Anomaly Diagnosis

Jian Wang , Yang Yang , Ziheng Pan , Xiliang Zhu , Yuhan Zhang , Yanfeng Zhou , Dong Ni This is my paper
classification cs.CV
keywords prenatal ultrasoundcongenital anomaliesweakly supervised classificationmultiple instance learningprototype learninganatomical plane identificationsparse selectionmedical image diagnosis
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

This paper argues that prenatal ultrasound anomaly diagnosis, trained only with case-level labels, improves when the model explicitly organizes images into latent anatomical planes and then selects the views most relevant to each anomaly. Its proposed framework, AnomExpert, uses K=30 learnable plane prototypes to group a case's unordered images through Sinkhorn-balanced soft assignments, then uses per-disease queries and a bias matrix to pick the top-4 planes before classification. On a 3,654-case, 24-center dataset of life-limiting congenital anomalies, the authors report 86.9% accuracy and 84.2% F1 with a vision-transformer backbone, outperforming nine multi-instance learning baselines. The paper concludes that modeling anatomical plane identification and disease-specific plane selection, not simply pooling all images, is what drives the improvement.

Core claim

The central claim is that modeling two clinical capabilities—anatomical plane identification and disease-specific plane selection—improves weakly supervised classification of prenatal ultrasound anomalies. The authors show that with only case-level diagnosis labels, learnable plane prototypes can organize unordered image sets into plane-level representations, and a disease-aware sparse selection mechanism can pick the diagnostically relevant planes for each anomaly. On their multi-center test set, this yields 86.9% accuracy, 84.2% F1, and 97.9% AUC with a ViT-small backbone, exceeding nine MIL baselines. Ablations indicate that the prototype loss, Sinkhorn balancing, top-k selection, and dis

What carries the argument

The central mechanism is a set of K=30 learnable plane prototypes (256-dimensional vectors) that act as latent anatomical-plane slots. Image features are softly assigned to prototypes with cosine similarity and Sinkhorn-balanced normalization, which encourages diverse prototype use; each prototype then pools the images assigned to it into a plane representation for a case. A set of learnable disease queries (one per class), plus a learnable disease-plane bias matrix, scores these plane representations, and the top-4 planes per disease are aggregated with softmax weights before a classifier. This machinery replaces the usual MIL pooling over all images with a two-stage 'identify planes, then

Load-bearing premise

The load-bearing premise is that the 30 learned prototypes align with real anatomical planes and that the top-4 selection picks the disease-relevant views—an assumption supported only by qualitative examples and ablations, not by clinician plane labels, so if the prototypes encode artifacts or label-correlated clusters instead, the claimed mechanism collapses.

What would settle it

Take a held-out set of images whose true anatomical plane is labeled by experienced sonographers, then inspect the prototype assignments produced by AnomExpert. If images sharing one prototype routinely span multiple different anatomical planes, or if the top-4 planes selected per anomaly do not overlap with the clinically relevant planes, the central interpretation fails. As a second check, replace the learned prototypes with fixed random prototypes and retrain; if the gap over MIL disappears, the plane-identification mechanism is not the source of the gains.

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

If this is right

  • With case-level labels only, high-accuracy prenatal anomaly diagnosis becomes feasible, so future datasets may not need time-consuming plane-level annotation.
  • Each component of the mechanism earns its place: ablations show that removing the prototype loss, replacing Sinkhorn with softmax, dense aggregation, and removing the disease bias each lower F1, supporting the two-stage design.
  • The framework scales down well: a 5.60M-parameter version still reaches 83.8% accuracy and 97.3% AUC, suggesting the plane-selection inductive bias, not raw capacity, drives the gain.
  • Because it consumes unordered image sets, the same identify-then-select architecture can be transferred to other multi-plane imaging tasks without requiring plane labels.
  • The class-wise selection heatmap indicates per-anomaly plane preferences, offering a route from black-box classification to a case-backed explanation of which anatomical views mattered.

Where Pith is reading between the lines

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

  • A natural next study would compare learned prototype assignments with sonographer-annotated plane labels; strong agreement would confirm the prototypes correspond to real anatomy, while disagreement would suggest the model may be exploiting non-anatomical correlations.
  • The top-4 plane choices could be turned into a clinical checklist: if the selected planes per disease match standard diagnostic views, AnomExpert could flag cases where a critical view is absent or poorly captured.
  • The same sparse plane-selection mechanism could be tested on other multi-view modalities such as fetal echocardiography or retinal imaging, where clinicians also reason over views rather than individual frames.
  • A direct causal test: freeze the learned plane prototypes and re-train only the selector, then remove the top-selected plane for one disease and check whether that disease's accuracy drops specifically; a drop would confirm that plane selection, not mere image pooling, carries the diagnostic signal.
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 / 8 minor

Summary. The paper introduces AnomExpert, a prototype-driven framework for weakly supervised prenatal ultrasound anomaly classification using only case-level labels. The method learns K=30 latent plane prototypes with Sinkhorn-balanced soft assignments (Sec. 2.1), forms case-level plane representations, and performs disease-aware top-k=4 plane selection via learnable disease queries and a bias matrix (Sec. 2.2). On a multi-center dataset of 3,654 cases with eight anomaly types and normal controls, AnomExpert with a ViT-small backbone reports 86.9% accuracy and 84.2% F1, outperforming nine MIL baselines in mean metrics. Component ablations show that the prototype loss, Sinkhorn balancing, top-k selection, and disease bias each contribute to performance.

Significance. If the central interpretation holds, AnomExpert is a meaningful step toward clinically useful multi-plane prenatal US diagnosis without plane-level annotation, a setting where manual plane labeling is costly. The paper's strengths include a large multi-center dataset, five-run statistics, a comparison against nine open-source baselines, component ablations, and release of the code. The conditional nature of the contribution, however, stems from the claim that the learned prototypes correspond to anatomical planes. The current evidence is only qualitative visual inspection and indirect ablation improvements, which do not establish semantic plane identity. With the interpretation either validated or appropriately softened to 'learned grouping,' the empirical method is sound and worth publishing.

major comments (3)
  1. [Sec. 2.1, Eq. (2-4); Fig. 3(a); Table 2] The central claim that prototypes correspond to anatomical planes is not supported by the evidence. Prototypes are learned with case-level diagnosis labels; without plane annotations, the clusters could correspond to image quality, fetal orientation, gestational age, or acquisition artifacts correlated with the diagnosis. The ablations in Table 2 show that the prototype mechanism helps classification, but this only demonstrates that the grouping is class-discriminative, not that it is anatomically meaningful. Please either (a) provide a quantitative evaluation against expert plane-level annotations on a held-out subset (e.g., prototype-to-plane agreement, retrieval of the clinically expected planes for each anomaly), or (b) revise the conclusion and abstract to avoid claiming 'anatomical plane identification' and instead refer to 'learned plane-like grouping.' The current title and abstr
  2. [Table 1 and Sec. 3.2] The claim that AnomExpert 'consistently outperforms' all nine baselines is not statistically established. No significance tests are reported. For example, F1 for AnomExpert (ViT-s) is 84.2±0.69 vs. 82.1±1.59 for MaxNetMIL, and accuracy is 86.9±0.72 vs. 85.5±1.06; the standard deviations overlap or nearly overlap. With five runs, paired tests (e.g., paired t-test or Wilcoxon signed-rank on per-run metrics, or a bootstrap over case-level predictions) should be reported to support the comparative claims. If the differences are not significant, the wording should be adjusted.
  3. [Sec. 3.1 and Fig. 4] The hyperparameters K, top-k, and lambda were each selected from five candidate values, but the paper does not specify whether this selection was performed on the validation set or the test set. If the test set was reused for hyperparameter tuning, the reported numbers are optimistically biased. Please state explicitly that the validation set was used (e.g., picking the values with best validation F1) and list the candidate grids. Also clarify whether the 'best performance' in Fig. 4 refers to validation or test metrics.
minor comments (8)
  1. [Sec. 2.2, Eq. (8)] The TopK notation is ambiguous for ties in relevance scores. Specify a deterministic tie-breaking rule (e.g., by index) or define the selection as a set-valued operation.
  2. [Sec. 2.2, Eq. (9)] The notation α_{c,Ω_c} = softmax(ũ_{c,Ω_c}/T) is imprecise; clarify that the softmax is over the selected indices and define α_{c,k} for k∈Ω_c explicitly.
  3. [Table 1] The abbreviation 'Anom.' is not defined; use 'AnomExpert' or introduce the abbreviation before the table. Also use consistent backbone names (ViT-t vs. ViT-tiny) throughout the text and table.
  4. [Fig. 3(b)] The caption mentions 'average selection probabilities' but the method section does not define these probabilities. Clarify whether these are the softmax weights from Eq. (9), selection frequencies from Eq. (8), or another aggregated quantity.
  5. [Fig. 4] The three subplots should have labeled axes and error bars; currently the reader cannot judge the scale of variation across hyperparameter values. Also report which metric is shown (F1-score is mentioned in the caption but not on axis labels).
  6. [Sec. 3.1, Datasets] The gestational age range and standard deviation appear inconsistent: '10 to 39 weeks (mean±std: 20.7±12.0)' would imply a very wide distribution; please verify these numbers or clarify whether the std is correct.
  7. [Eq. (4)] The constant δ is introduced as 'a small constant for numerical stability' but its value is never given. State a concrete value (e.g., 1e-6) or note that it is set to a standard epsilon.
  8. [References] Reference formatting is inconsistent (e.g., [7] NIPS volume/year, [2] 'Nature medicine' capitalization, [19] WACV year). Please unify according to the journal style.

Circularity Check

0 steps flagged

No circularity: AnomExpert's results are held-out empirical evaluations; prototype-plane interpretation is an unverified assumption, not a circular derivation.

full rationale

The paper's derivation chain is not circular. The model is a standard end-to-end weakly supervised classifier: image features (Eq. 1) are assigned to learnable prototypes via Sinkhorn normalization (Eq. 2), aggregated into plane representations (Eq. 4), selected by disease queries (Eqs. 6-9), and trained with case-level cross-entropy plus a prototype loss (Eqs. 11-12). No equation reduces to a fitted constant or to the reported test numbers, and the test set is held out with stratified case-level splitting. The claimed performance (86.9% accuracy, 84.2% F1) is an empirical outcome, not an algebraic consequence of the method's definitions. The ablations in Table 2 demonstrate component contributions empirically rather than defining the result. The only apparent self-citation is Mamba2MIL (Ref. 20), used as a comparison baseline; it is not load-bearing for any claim. The substantive weakness—that the learned prototypes are asserted to correspond to anatomical planes without clinician plane annotations—is an interpretability/validation gap, not a circular-equation or self-referential-derivation issue. Therefore the appropriate verdict is no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The method is a supervised classifier; its free parameters are the usual learned weights plus hand-set/tuned hyperparameters. The two conceptual commitments—that prototypes approximate anatomical planes and that sparse plane selection is disease-specific—are model-internal and are not independently validated against plane-level labels.

free parameters (7)
  • Number of plane prototypes K = 30
    Selected by sweeping five candidate values (Fig 4); best at 30.
  • Top-k selected planes = 4
    Selected from candidate values (Fig 4); best at 4.
  • Prototype loss weight lambda = 0.1
    Selected from candidate values (Fig 4); best at 0.1.
  • Sinkhorn temperature epsilon_s = 0.05
    Hand-set control for assignment sharpness in Eq. 2.
  • Selection temperature T = 0.07
    Hand-set temperature for softmax selection in Eq. 9.
  • Learnable temperature tau = learned
    Temperature in Eq. 3 optimized during training.
  • Prototype, query, bias and classifier parameters = trained
    P, Q, B, classifier, and backbone weights are learned end-to-end from training data.
axioms (5)
  • standard math Sinkhorn-Knopp balancing yields differentiable doubly-stochastic assignments that prevent prototype collapse.
    Used in Eq. 2; assumes the standard optimal-transport result from Cuturi [7].
  • domain assumption Each case consists of unordered images with a single primary diagnosis label.
    The dataset design and the cross-entropy loss in Eq. 11 rely on one case-level label per case; no multi-label or plane-level supervision is used.
  • ad hoc to paper Thirty learnable prototypes can capture the anatomical-plane structure relevant to all classes.
    No theoretical or anatomical justification is given for K=30; it is an empirical architecture choice (Fig 4).
  • domain assumption Specialist-confirmed diagnoses are reliable ground truth.
    Labels were confirmed by experienced fetal US specialists; no inter-observer agreement or adjudication protocol is reported (Sec 3.1).
  • domain assumption ImageNet-pretrained ViT features transfer to prenatal ultrasound.
    Both backbones are ImageNet-pretrained and fine-tuned on ultrasound; no ultrasound-specific pretraining is used (Sec 3.1).
invented entities (2)
  • Learnable plane prototypes P_k no independent evidence
    purpose: Latent categories representing anatomical planes; images are soft-assigned to prototypes to form plane-level representations.
    No plane labels or clinician annotations verify that the prototypes correspond to actual anatomical planes; only qualitative Figure 3a.
  • Disease queries Q_c and bias matrix B no independent evidence
    purpose: Estimate per-disease relevance of each plane and select top-k planes before classification.
    Learned end-to-end; the claimed disease-plane associations are not grounded in an external anatomical atlas or prior clinical knowledge.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AnomExpert: Identifying and Selecting Anatomical Planes for Prenatal Ultrasound Anomaly Diagnosis." pith.science (2026). https://pith.science/paper/BD33HFHT

@misc{pith2026260713409,
  author       = {Pith},
  title        = {Pith review of: AnomExpert: Identifying and Selecting Anatomical Planes for Prenatal Ultrasound Anomaly Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BD33HFHT}},
  note         = {Machine review of arXiv:2607.13409}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Life-limiting congenital anomalies require accurate prenatal diagnosis for appropriate clinical decision-making. Prenatal ultrasound (US) examinations involve multiple anatomical planes, and diagnosis depends on identifying anatomical planes and selecting diagnostically relevant planes for each anomaly. Existing automated methods either rely on plane-level annotations or aggregate heterogeneous images without explicitly modeling these diagnostic capabilities. We propose AnomExpert, a prototype-driven framework for prenatal US anomaly diagnosis using only case-level supervision. AnomExpert introduces learnable plane prototypes to organize unordered images into latent representations corresponding to anatomical planes without requiring plane annotations. A disease-aware sparse selection mechanism further selects diagnostically relevant planes for each anomaly. Experiments on a multi-center dataset of 3,654 cases show that AnomExpert consistently outperforms nine representative multi-instance learning methods. Using a ViT-small backbone, it achieves 86.9% accuracy and 84.2% F1-score while maintaining parameter efficiency. These findings indicate that modeling anatomical plane identification and disease-specific plane selection improves weakly supervised multi-plane prenatal US anomaly classification. The code is available at https://github.com/TIanCat/AnomExpert.

Figures

Figures reproduced from arXiv: 2607.13409 by Dong Ni, Jian Wang, Xiliang Zhu, Yanfeng Zhou, Yang Yang, Yuhan Zhang, Ziheng Pan.

Figure 1
Figure 1. Figure 1: Prenatal Ultrasound Anomaly Diagnosis [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed AnomExpert [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative visualization. (a) Test images assigned to selected plane prototypes. (b) Class-wise relevance heatmap showing disease-plane associations. resized to 224×224. Additional data augmentations included horizontal flipping, color jittering, affine transformations, and random grayscale conversion. Evaluation Protocol We compared our method with nine representative multi￾instance learning (MIL) approa… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on key hyperparameters. F1-score versus plane prototype num￾ber K (left), top-k selection (middle), and loss weight λ (right) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

20 extracted references · 1 linked inside Pith

  1. [1]

    New England Journal of Medicine364(11), 993–1004 (2011)

    Adzick, N.S., Thom, E.A., Spong, C.Y., Brock III, J.W., Burrows, P.K., Johnson, M.P., Howell, L.J., Farrell, J.A., Dabrowiak, M.E., Sutton, L.N., et al.: A random- ized trial of prenatal versus postnatal repair of myelomeningocele. New England Journal of Medicine364(11), 993–1004 (2011)

    2011

  2. [2]

    Nature medicine27(5), 882–891 (2021)

    Arnaout, R., Curran, L., Zhao, Y., Levine, J.C., Chinn, E., Moon-Grady, A.J.: An ensemble of neural networks provides expert-level prenatal detection of complex congenital heart disease. Nature medicine27(5), 882–891 (2021)

    2021

  3. [3]

    IEEE Access12, 78409–78422 (2024)

    Barbosa, D., Ferreira, M., Junior, G.B., Salgado, M., Cunha, A.: Multiple instance learning in medical images: a systematic review. IEEE Access12, 78409–78422 (2024)

    2024

  4. [4]

    In: Seminars in Fetal and Neonatal Medicine

    Breeze, A.C., Lees, C.C.: Antenatal diagnosis and management of life-limiting con- ditions. In: Seminars in Fetal and Neonatal Medicine. vol. 18, pp. 68–75. Elsevier (2013)

    2013

  5. [5]

    International Journal of Gynecology & Obstetrics102(3), 304–308 (2008) 10 J

    Cook, R.J., Erdman, J.N., Hevia, M., Dickens, B.M.: Prenatal management of anencephaly. International Journal of Gynecology & Obstetrics102(3), 304–308 (2008) 10 J. Wang et al

    2008

  6. [6]

    Fetal diagnosis and therapy50(6), 480–490 (2023)

    Coronado-Gutierrez, D., Eixarch, E., Monterde, E., Matas, I., Traversi, P., Grat- acos, E., Bonet-Carne, E., Burgos-Artizzu, X.P.: Automatic deep learning-based pipelineforautomaticdelineationandmeasurementoffetalbrainstructuresinrou- tine mid-trimester ultrasound images. Fetal diagnosis and therapy50(6), 480–490 (2023)

    2023

  7. [7]

    Ad- vances in neural information processing systems26(2013)

    Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Ad- vances in neural information processing systems26(2013)

    2013

  8. [8]

    IEEE Journal of Biomedical and Health Informatics27(10), 5023–5031 (2022)

    Guo,J.,Tan,G.,Wu,F.,Wen,H.,Li,K.:Fetalultrasoundstandardplanedetection with coarse-to-fine multi-task learning. IEEE Journal of Biomedical and Health Informatics27(10), 5023–5031 (2022)

    2022

  9. [9]

    In: Machine Learning for Healthcare Conference

    Huang, Z., Wessler, B.S., Hughes, M.C.: Detecting heart disease from multi-view ultrasound images via supervised attention multiple instance learning. In: Machine Learning for Healthcare Conference. pp. 285–307. PMLR (2023)

    2023

  10. [10]

    In: International conference on machine learning

    Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: International conference on machine learning. pp. 2127–2136. PMLR (2018)

    2018

  11. [11]

    Ilse, M., Tomczak, J.M., Welling, M.: Deep multiple instance learning for digital histopathology.In:HandbookofMedicalImageComputingandComputerAssisted Intervention, pp. 521–546. Elsevier (2020)

    2020

  12. [12]

    Ad- vances in Neural Information Processing Systems35, 20689–20702 (2022)

    Javed, S.A., Juyal, D., Padigela, H., Taylor-Weiner, A., Yu, L., Prakash, A.: Ad- ditive mil: Intrinsically interpretable multiple instance learning for pathology. Ad- vances in Neural Information Processing Systems35, 20689–20702 (2022)

    2022

  13. [13]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, B., Li, Y., Eliceiri, K.W.: Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2021)

    2021

  14. [14]

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

    Liang, H., Xu, J., Zhang, Y., Huang, Y., Zhang, Y., Yang, X., Li, R., Deng, X., Liu, Y., Tao, G., et al.: Medical-knowledge driven multiple instance learning for classifying severe abdominal anomalies on prenatal ultrasound. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 344–354. Springer (2025)

    2025

  15. [15]

    Circulation134(17), 1265– 1279 (2016)

    Ohye, R.G., Schranz, D., D’Udekem, Y.: Current therapy for hypoplastic left heart syndrome and related single ventricle lesions. Circulation134(17), 1265– 1279 (2016)

    2016

  16. [16]

    Ultrasound in Obstetrics and Gynecology59(6), 840–856 (2022)

    Salomon, L., Alfirevic, Z., Berghella, V., Bilardo, C., Chalouhi, G., Costa, F.D.S., Hernandez-Andrade, E., Malinger, G., Munoz, H., Paladini, D., et al.: Isuog prac- tice guidelines (updated): performance of the routine mid-trimester fetal ultra- sound scan. Ultrasound in Obstetrics and Gynecology59(6), 840–856 (2022)

    2022

  17. [17]

    Ultrasound in obstetrics & gynecology39(3), 266–273 (2012)

    Sarris, I., Ioannou, C., Chamberlain, P., Ohuma, E., Roseman, F., Hoch, L., Alt- man, D., Papageorghiou, A., Fetal, I., for the 21st Century (INTERGROWTH- 21st),N.G.C.:Intra-andinterobservervariabilityinfetalultrasoundmeasurements. Ultrasound in obstetrics & gynecology39(3), 266–273 (2012)

    2012

  18. [18]

    Advances in neural information processing systems34, 2136–2147 (2021)

    Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Trans- former based correlated multiple instance learning for whole slide image classifica- tion. Advances in neural information processing systems34, 2136–2147 (2021)

    2021

  19. [19]

    In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Shiku, K., Nishimura, K., Suehiro, D., Tanaka, K., Bise, R.: Ordinal multiple- instance learning for ulcerative colitis severity estimation with selective aggregated transformer. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4290–4299. IEEE (2025)

    2025

  20. [20]

    arXiv preprint arXiv:2408.15032 (2024)

    Zhang, Y., Zhang, X., Wang, J., Yang, Y., Peng, T., Tong, C.: Mamba2mil: State space duality based multiple instance learning for computational pathology. arXiv preprint arXiv:2408.15032 (2024)

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.