Pith. sign in

REVIEW 4 major objections 4 minor 38 references

MoPET: Parameter-Efficient Mixture-of-Experts for Unified Medical Image Classification

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A single network of 32 routed low-rank experts outperforms isolated per-task adapters on four medical image datasets, and co-training data-constrained targets with auxiliary datasets lifts their accuracy.

desk verdict A credible empirical paper whose unified-model result is plausible but underproved, and whose booster claim is confounded by added data—exactly as the authors themselves concede. read the letter →

arxiv 2607.29462 v2 pith:JYFHTP7M submitted 2026-07-31 eess.IV cs.CVcs.LG

Sebastian Doerrich , Daniel Würtinger , Francesco Di Salvo , Shyam Nandan Rai , Christian Ledig This is my paper
classification eess.IVcs.CVcs.LG
keywords medicalimageclassificationmixture-of-expertsparameter-efficientfine-tuningLoRABOFTsparseroutingMedMNISTtransferlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a single frozen foundation model can serve multiple medical image classification tasks at once without the usual negative transfer, using a learned sparse mixture of low-rank adapters (MoPET). It first shows that parameter-efficient fine-tuning (LoRA, BOFT) beats full end-to-end updates across twelve MedMNIST datasets. It then shows that one MoPET model consolidating four heterogeneous datasets outperforms the best isolated per-task adapter on average, and that co-training data-constrained targets with auxiliary datasets lifts their accuracy. If true, this points to a deployable alternative to the current practice of training and storing a separate adapter per clinical task, as well as a possible way to improve small-data medical targets.

What carries the argument

The central mechanism is the learned sparse router over a heterogeneous pool of low-rank experts. The router computes routing scores for each input token, keeps the top-k (k=12) experts, and applies a softmax over those scores to weight the experts' outputs, which are added to the frozen projection path. This lets the model share capacity across domains while isolating conflicting gradients; a load-balancing loss prevents collapse to a few experts. Dynamic multi-domain sampling caps each epoch to the smallest dataset's sample count, and per-dataset heads handle disjoint label spaces.

What would settle it

A concrete way to test the booster claim: train MoPET on BreastMNIST with its chosen auxiliary pool (Blood, Derma, Path) and again with three randomly selected MedMNIST datasets matched in total sample size; if the accuracy gain over the isolated BOFT baseline (90.38% -> 92.95%) does not reproduce or exceed the random-pool version, the semantically driven transfer claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that a mixture-of-experts built entirely from parameter-efficient adapters can unify multiple medical image classification tasks in one network and beat the per-task adapters it consolidates. Specifically, with a frozen DINOv3 backbone, 32 low-rank experts per block (20 LoRA, 12 BOFT), sparse top-12 routing, per-dataset heads, and dynamic sampling, MoPET raises average accuracy on Blood-, Breast-, Derma-, and PathMNIST from 92.83% (best isolated BOFT) to 93.46%. A second result is that co-training a small target dataset (Breast, Retina, or Derma) with hand-selected auxiliary datasets raises average target accuracy from 81.58% to 83.58%, with the largest gains on the smal

Load-bearing premise

The booster results assume that the hand-selected auxiliary pools help because of semantically meaningful feature sharing, not merely because they add more training data, and the paper ran no control with random or unrelated auxiliary datasets.

Editorial extensions

If this is right

  • A single MoPET model can replace a collection of isolated adapters for several imaging modalities, reducing deployment complexity and memory footprint while slightly improving average accuracy.
  • Co-training with auxiliary data can improve accuracy on data-constrained clinical targets, with gains largest on the smallest datasets (Breast, Retina), suggesting a practical recipe for rare-disease classification.
  • The rank ordering PEFT > full fine-tuning, and LoRA/BOFT > FourierFT/AdaptFormer, gives a data point for choosing adaptation methods on medical imaging benchmarks.
  • The hybrid LoRA/BOFT expert pool outperforms any single adapter family, suggesting that complementary low-rank methods can be combined to advantage.
  • The recipe is forward-compatible: the paper's scaling convention suggests larger expert pools could extend the approach beyond four datasets.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The booster effect is not yet causally separated from the sheer addition of training data; the paper's own limitation note implies a control with randomly selected auxiliary datasets of matched size would settle whether semantic transfer or data volume drives the 83.58% result.
  • If the router's assignments were visualized per dataset, one could test whether experts actually specialize by modality/anatomy; the paper leaves this unexamined, so the 'domain-specialized pathways' story is an untested interpretation.
  • A natural testable extension is to apply MoPET to 3D medical volumes (e.g., MedMNIST3D) to see whether routing-based consolidation holds beyond 2D slices, where gradient conflicts may be stronger.
  • The hand-selected booster pools were not swept systematically; a broader combinatorial search over auxiliary subsets could reveal whether small pools, rather than general co-training, are what helps.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes MoPET, a mixture-of-experts architecture that couples a learned sparse router with a heterogeneous pool of LoRA and BOFT experts injected into a frozen DINOv3 backbone, aiming to unify multiple medical image classification tasks in a single network. The authors claim three main results: (1) across 12 MedMNIST datasets, parameter-efficient fine-tuning (PEFT), especially LoRA and BOFT, outperforms full end-to-end fine-tuning; (2) a single MoPET model trained on four heterogeneous datasets improves average test accuracy over the best isolated PEFT adapter (93.46% vs 92.83%); (3) co-training with hand-selected auxiliary datasets boosts accuracy on three data-constrained target datasets (average 83.58% vs 81.58% for the strongest isolated adapter). The paper releases source code and includes a supplementary document with extended comparisons, statistical tests, and robustness analyses.

Significance. If the claims hold, MoPET is a practically useful contribution to parameter-efficient unified medical image classification: it consolidates multiple task-specific adapters into one network while using only 8.7% of the backbone parameters, and it identifies a potentially interesting auxiliary-data boosting effect. The paper is commendable for releasing code, providing detailed supplementary material, and applying non-parametric significance tests to the PEFT-vs-end-to-end comparison. However, the headline cross-dataset and booster results are presented without variance or significance testing, and the booster experiment is confounded by simultaneous changes in training-set size and composition, so the central claims are not yet fully supported.

major comments (4)
  1. [§3.3, Table 3] The booster claim is confounded. In Table 3, MoPET is trained on the target plus a hand-selected auxiliary pool, while the isolated adapters are trained on the target alone. Training-set size and composition change simultaneously, so the gains (e.g., Breast 90.38→92.95, Retina 67.42→68.83) could be driven by additional data quantity or regularization rather than semantically driven cross-domain transfer. The Limitations section explicitly concedes that no control with randomly selected or unrelated auxiliary data was run. Because the abstract and contributions advertise this 'cross-domain training dynamic' as a novel finding, a control condition with equal-size random/unrelated auxiliary pools is required to support the causal interpretation.
  2. [§3.2–§3.3, Tables 2 and 3] Tables 2 and 3 report only means over three seeds, with no standard deviations, confidence intervals, or significance tests. The headline differences are small—in Table 2 the average gain is 0.63 percentage points (93.46 vs 92.83)—and per-dataset differences are often below one point. Without seed-level data or a paired significance test (e.g., permutation or Wilcoxon signed-rank on the three seeds, or per-dataset repeated runs), these improvements are not distinguishable from sampling noise. Please report individual seed results and perform a paired test for the MoPET vs best-isolated-adapter comparisons.
  3. [§2.3; Limitations] The architecture's central hyperparameters—K=32 experts per block, top-k=12 active experts, the 20-LoRA/12-BOFT composition, and the load-balancing weight λ=0.05—are fixed without ablation. The Limitations section itself acknowledges that K, top-k, and the expert ratio are 'likely dataset dependent' and warrant a more rigorous ablation. Since the paper's central contribution is the MoPET design, an ablation over at least K, top-k, and the LoRA-to-BOFT ratio is needed to establish that the observed gains over isolated adapters are not artifacts of a particular hand-tuned configuration.
  4. [§3.1] The end-to-end baselines for CLIP and DINO, and apparently the DINOv3 training procedure, are 'adapted from [11]' rather than re-run in this paper. If the PEFT numbers are produced with different seeds, resolution, or hyperparameters than the numbers taken from [11], then Table 1's comparison is not properly controlled. Please clarify exactly which numbers are newly produced, which are reused, and confirm that the training protocols match (including early stopping, resolution, and seed counts); if they do not, the end-to-end baselines should be re-run under the same conditions.
minor comments (4)
  1. [Eq. (2)] The load-balancing loss is written as L_load = K * (1/K) Σ f_i p_i, which algebraically reduces to Σ f_i p_i. The factor K and any additional scaling coefficient (as used in DeepSeekMoE) should be stated explicitly, and the final objective's λ should be clearly defined in Eq. (2) or the surrounding text.
  2. [Table 2 caption] The caption says MoPET entries are shown in bold when they rank among the top three, but the table formatting is not visible in the manuscript text. Please make the bold/underline convention unambiguous for all methods.
  3. [§2.1] The text says the routing stage operates 'without modality labels', but then a dataset router uses the domain identifier d_id. Please clarify that d_id is used only for the final classification-head selection and not for expert routing, to avoid confusion.
  4. [§3.2] The selection of the four datasets in Table 2 and the three booster pools in Table 3 is described as 'hand-selected' or chosen to 'span modalities and scales'. Please state the explicit selection criteria and discuss whether the reported results are sensitive to this choice, especially since a single dataset pair (e.g., BloodMNIST) can change the average by a small margin.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation-level circularity; the paper is an empirical benchmark study, and its only self-citation is a baseline source. The booster confound is explicitly acknowledged as a limitation, not a circular step.

full rationale

The paper's claims are empirical, supported by held-out test splits and multi-seed means, rather than by a derivation chain where an output equals an input by construction. The only self-citation appears in Section 3.1: "The end-to-end results for DINO and CLIP, as well as the DINOv3 training procedure, are adapted from [11]." This is a baseline taken from the authors' prior published work; the PEFT numbers are trained and evaluated in the present paper, and the comparison is an external benchmark, so no equation or fitted parameter reduces to an input. The booster experiment (Section 3.3, Table 3) changes training-set size and composition simultaneously, and the Limitations section concedes that "control conditions using randomly selected or unrelated auxiliary data, would be needed to separate semantically driven transfer from the effect of added training data." This is a causal-identification confound, not circularity: the result is not equivalent to its own input by construction. Design choices such as K=32 experts, top-k=12, and the LoRA/BOFT ratio are motivated by earlier isolated experiments, not fitted to the test set. No circular step meets the quoted-reduction test, so the score is low.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No new physical or ontological entities are introduced; LoRA/BOFT experts, routers, and heads are standard constructs. The listed free parameters are hand-picked hyperparameters that the central results depend on, and the axioms are domain assumptions about transfer, routing behavior, and the booster effect.

free parameters (7)
  • K (expert count per block) = 32
    Hand-selected following scaling conventions; paper's limitations say it is likely dataset dependent and no ablation was run.
  • top-k active experts = 12
    Sparse routing threshold chosen without ablation; affects capacity and gradient isolation.
  • LoRA-to-BOFT expert ratio = 20 LoRA : 12 BOFT
    Chosen from observed complementarity of isolated PEFT adapters; no ablation within MoPET.
  • LoRA rank and scaling = r=8, alpha=8
    Standard LoRA configuration; no sensitivity analysis reported.
  • BOFT block size and butterfly factor = block=8, butterfly=1
    Default BOFT settings; no ablation reported.
  • Load-balancing loss weight lambda = 0.05
    Hyperparameter balancing L_CE and L_load; no ablation reported.
  • Training schedule hyperparameters = 75 epochs, batch 128, LR 1e-3, patience 10
    Standard AdamW/cosine settings; baseline LR reduced to 1e-4. These affect all reported numbers.
assumptions (5)
  • domain assumption DINOv3 frozen representations transfer sufficiently to the evaluated medical tasks.
    All MoPET results rely on this; no evidence outside MedMNIST is provided.
  • domain assumption The DeepSeekMoE load-balancing loss improves expert utilization without hurting task accuracy.
    Adopted from [9] and used without validation on medical classification.
  • domain assumption Negative transfer in this setting is caused by gradient conflict between domains and is mitigated by token-level sparse routing.
    Motivates the architecture; routing patterns are not analyzed, as the authors acknowledge.
  • ad hoc to paper Hand-selected auxiliary dataset pools improve targets through semantic feature sharing rather than simply providing more training data.
    The booster claim depends on this; the paper admits no random/unrelated auxiliary control was run.
  • domain assumption MedMNIST datasets are representative of 'profound clinical heterogeneity'.
    Used to generalize from benchmark results to broader clinical deployment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoPET: Parameter-Efficient Mixture-of-Experts for Unified Medical Image Classification." pith.science (2026). https://pith.science/paper/JYFHTP7M

@misc{pith2026260729462,
  author       = {Pith},
  title        = {Pith review of: MoPET: Parameter-Efficient Mixture-of-Experts for Unified Medical Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYFHTP7M}},
  note         = {Machine review of arXiv:2607.29462}
}
read the original abstract

Adapting deep learning models to profound clinical heterogeneity typically relies on parameter-efficient fine-tuning (PEFT) to avoid the severe overfitting associated with full end-to-end network updates. Although PEFT successfully navigates limited data scenarios, it inherently forces the training of a separate, isolated adapter for every specific diagnostic task. Consolidating these isolated adapters into a single generalist network risks negative transfer, as optimization gradients from conflicting visual domains interfere. To address this, we propose MoPET, a mixture-of-experts (MoE) method that uses a learned sparse router to direct each input through a small subset of low-rank PEFT experts injected into a frozen foundation model, sharing capacity across datasets while limiting cross-domain gradient conflict. Through selected evaluations on the MedMNIST benchmark, we first establish that PEFT outperforms full network updates, improving average accuracy from 86.50% to 88.97%. We then show that a single MoPET model consolidates four heterogeneous datasets into one network, improving average accuracy over the best isolated PEFT adapters (93.46% versus 92.83%). Finally, we show that co-training with auxiliary datasets improves accuracy on data-constrained clinical targets, raising average target accuracy over the strongest isolated adapter from 81.58% to 83.58%. Our source code is publicly available at https://github.com/sdoerrich97/mopet.

Figures

Figures reproduced from arXiv: 2607.29462 by the authors.

Figure 1
Figure 1. Left: Standard Parameter-Efficient Fine-Tuning (PEFT) requires independent, disconnected models for distinct medical modalities. Right: MoPET dynamically routes diverse anatomical inputs through specialized low-rank experts within a single unified architecture, preventing gradient collision and boosting overall accuracy. in narrow anatomical regions, demanding necessary adaptation for broader clini￾cal application. … view at source ↗
Figure 2
Figure 2. Overview of MoPET. First, a dynamic sampling protocol extracts balanced batches of size dmin from each dataset (A). Second, during the forward pass, an internal router activates the top-k parameter-efficient (PEFT) experts to process the input representation x alongside the frozen pretrained weights, producing the updated token representation z (B). Finally, a dataset router utilizes the domain identifier did to dir… view at source ↗

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 1 linked inside Pith

  1. [11]

    Scientific Reports15(1), 7669 (Mar 2025)

    Doerrich, S., Di Salvo, F., Brockmann, J., Ledig, C.: Rethinking model prototyping through the MedMNIST+ dataset collection. Scientific Reports15(1), 7669 (Mar 2025)

    2025

  2. [1]

    Data in Brief30, 105474 (2020)

    Acevedo, A., Merino, A., Alférez, S., Ángel Molina, Boldú, L., Rodellar, J.: A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief30, 105474 (2020)

    2020

  3. [2]

    Data in Brief28, 104863 (2020)

    Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in Brief28, 104863 (2020)

    2020

  4. [3]

    Medical Image Analysis84, 102680 (2023)

    Bilic, P., Christ, P., Li, H.B., Vorontsov, E., Ben-Cohen, A., et al.: The liver tumor segmentation benchmark (lits). Medical Image Analysis84, 102680 (2023)

    2023

  5. [4]

    2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp

    Caron, M., Touvron, H., Misra, I., J’egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 9630–9640 (2021)

    2021

  6. [5]

    Nature medicine30(3), 850–862 (Mar 2024)

    Chen, R.J., Ding, T., Lu, M.Y., et al.: Towards a General-Purpose Foundation Model for Computational Pathology. Nature medicine30(3), 850–862 (Mar 2024)

    2024

  7. [6]

    Advances in Neural Information Processing Systems (NeurIPS)35(2022)

    Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems (NeurIPS)35(2022)

    2022

  8. [7]

    Chopra, S., Sanchez-Rodriguez, G., Mao, L., Feola, A.J., Li, J., Kira, Z.: MedMoE: Modality-SpecializedMixtureofExpertsforMedicalVision-LanguageUnderstand- ing (Jun 2025)

    2025

Show all 38 references
  1. [8]

    In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)

    Codella, N.C.F., Gutman, D., Celebi, M.E., et al.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomed- ical imaging (isbi), hosted by the international skin imaging collaboration (isic). In: 2018 IEEE 15th International Sy...

    2017

  2. [9]

    Doerrich et al

    Dai, D., Deng, C., Zhao, C., et al.: DeepSeekMoE: Towards Ultimate Expert Spe- cialization in Mixture-of-Experts Language Models (Jan 2024) 10 S. Doerrich et al

    2024

  3. [10]

    arXiv preprint arXiv:2406.17536 (2024)

    Di Salvo, F., Doerrich, S., Ledig, C.: MedMNIST-C: Comprehensive benchmark and improved classifier robustness by simulating realistic image corruptions. arXiv preprint arXiv:2406.17536 (2024)

    2024 arXiv

  4. [12]

    ICML’24, JMLR.org (2024)

    Gao, Z., Wang, Q., Chen, A., Liu, Z., Wu, B., Chen, L., Li, J.: Parameter-efficient fine-tuningwithdiscretefouriertransform.In:Proceedingsofthe41stInternational Conference on Machine Learning. ICML’24, JMLR.org (2024)

    2024

  5. [13]

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-Rank Adaptation of Large Language Models (Oct 2021)

    2021

  6. [14]

    Neural Computation3(1), 79–87 (Feb 1991)

    Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive Mixtures of Local Experts. Neural Computation3(1), 79–87 (Feb 1991)

    1991

  7. [15]

    Jiang, S., Zheng, T., Zhang, Y., Jin, Y., Yuan, L., Liu, Z.: Med-MoE: Mixture of Domain-Specific Experts for Lightweight Medical Vision-Language Models (Sep 2024)

    2024

  8. [16]

    In: proceedings of Medical Image Com- puting and Computer Assisted Intervention – MICCAI 2024

    Jiang, Y., Shen, Y.: M4oE: A Foundation Model for Medical Multimodal Image Segmentation with Mixture of Experts . In: proceedings of Medical Image Com- puting and Computer Assisted Intervention – MICCAI 2024. vol. LNCS 15012. Springer Nature Switzerland (2024)

    2024

  9. [17]

    PLOS Medicine16, e1002730 (2019)

    Kather, J.N., Krisam, J., Charoentong, P., Luedde, T., Herpel, E., et al.: Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLOS Medicine16, e1002730 (2019)

    2019

  10. [18]

    Cell 172, 1122–1131.e9 (2018)

    Kermany, D.S., Goldbaum, M., Cai, W., Valentim, C.C., Liang, H., et al.: Identi- fying medical diagnoses and treatable diseases by image-based deep learning. Cell 172, 1122–1131.e9 (2018)

    2018

  11. [19]

    In: International Conference on Learning Representations (ICLR) (2023)

    Komatsuzaki,A.,Puigcerver,J.,Lee-Thorp,J.,Ruiz,C.R.,Mustafa,B.,Ainslie,J., Tay, Y., Dehghani, M., Houlsby, N.: Sparse upcycling: Training mixture-of-experts from dense checkpoints. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  12. [20]

    Patterns3, 100512 (2022)

    Liu, R., Wang, X., Wu, Q., Dai, L., Fang, X., et al.: Deepdrid: Diabetic retinopa- thy—grading and image quality estimation challenge. Patterns3, 100512 (2022)

    2022

  13. [21]

    In: ICLR (2024)

    Liu, W., Qiu, Z., Feng, Y., Xiu, Y., Xue, Y., et al.: Parameter-efficient orthogonal finetuning via butterfly factorization. In: ICLR (2024)

    2024

  14. [22]

    Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts (2017)

    2017

  15. [23]

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019)

    2019

  16. [24]

    OpenAI, Agarwal, S., Ahmad, L., et al.: Gpt-oss-120b & gpt-oss-20b Model Card (Aug 2025)

    2025

  17. [25]

    Pan,D.,Li,B.,Zheng,Y.,Ma,J.,Fei,V.:TheRiseofSparseMixture-of-Experts:A Survey from Algorithmic Foundations to Decentralized Architectures and Vertical Domain Applications (Feb 2026)

    2026

  18. [26]

    In: Meila, M., Zhang, T

    Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 874...

    2021

  19. [27]

    Ruder, S.: An Overview of Multi-Task Learning in Deep Neural Networks (Jun 2017)

    2017

  20. [28]

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., et al.: DINOv3 (Aug 2025) Parameter-Efficient MoE for Unified Medical Image Classification 11

    2025

  21. [29]

    Nature Biomedical Engineering6(12), 1399–1406 (Sep 2022)

    Tiu, E., Talius, E., Patel, P., Langlotz, C.P., Ng, A.Y., Rajpurkar, P.: Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning. Nature Biomedical Engineering6(12), 1399–1406 (Sep 2022)

    2022

  22. [30]

    Scientific Data 2018 5:15, 1–9 (2018)

    Tschandl, P., Rosendahl, C., Kittler, H.: The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data 2018 5:15, 1–9 (2018)

    2018

  23. [31]

    2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised clas- sification and localization of common thorax diseases. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CV...

    2017

  24. [32]

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: MedCLIP: Contrastive Learning from Unpaired Medical Images and Text (Oct 2022)

    2022

  25. [33]

    Wightman, R.: Pytorch image models.https://github.com/rwightman/ pytorch-image-models(2019)

    2019

  26. [34]

    IEEE Transactions on Medical Imaging 38(8), 1885–1898 (2019)

    Xu, X., Zhou, F., Liu, B., Fu, D., Bai, X.: Efficient multiple organ localization in ct image using 3d region proposal network. IEEE Transactions on Medical Imaging 38(8), 1885–1898 (2019)

    2019

  27. [35]

    Scientific Data10(1), 41 (Jan 2023)

    Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., Ni, B.: MedM- NIST v2 – A large-scale lightweight benchmark for 2D and 3D biomedical image classification. Scientific Data10(1), 41 (Jan 2023)

    2023

  28. [36]

    Yang, J., Shi, R., Wei, D., Liu, Z., et al.: MedMNIST+: 18×standardized datasets for 2d and 3d biomedical image classification with multiple size options: 28 (mnist- like), 64, 128, and 224 (2024)

    2024

  29. [37]

    In: International Conference on Learning Representations (ICLR) (2024)

    Zadouri, T., Üstün, A., Ahmadian, A., Ermiş, B., Locatelli, A., Hooker, S.: Pushing mixture of experts to the limit: Extremely parameter efficient MoE for instruction tuning. In: International Conference on Learning Representations (ICLR) (2024)

    2024

  30. [38]

    Isolated PEFT

    Zhang, S., Xu, Y., Usuyama, N., et al.: BiomedCLIP: A multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs (Jan 2025) Parameter-Efficient MoE for Unified Medical Image Classification S1 Overview of Supplementary Material Thissupple...

    2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.