Foundation Models as Class-Incremental Learners for Dermatological Image Classification

Jamil Fayyad; Mohamed Elkhayat; Mohamed Mahmoud; Nourhan Bayasi

REVIEW 4 major objections 5 minor 1 cited by

Foundation Models as Class-Incremental Learners for Dermatological Image Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Frozen foundation models outperform specialized continual-learning methods for dermatology image classification, achieving zero forgetting.

desk verdict Useful baseline study, but the derm-FM results are compromised by a plausible pretraining/test overlap that the authors never address. read the letter →

arxiv 2507.14050 v1 pith:ONP4CGGN submitted 2025-07-18 cs.CV

Mohamed Elkhayat , Mohamed Mahmoud , Jamil Fayyad , Nourhan Bayasi This is my paper · ORCID

classification cs.CV

keywords class-incrementallearningcontinualfoundationmodelsdermatologyskinlesionclassificationfrozenbackbonenearestmeanclassifiercatastrophicforgetting

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

This paper tries to establish that a frozen foundation model, one whose weights are never updated, can serve as a complete class-incremental learner for dermatology image classification. The authors show that training only a small MLP on top of the frozen embeddings, task by task, beats regularization, replay, and architecture-based continual learning methods across three skin-lesion benchmarks, and even exceeds a single-task upper bound on HAM10000. They also demonstrate a no-training variant, a nearest-mean classifier built from class-mean embeddings, which is competitive when the embeddings are projected and normalized. If the claim holds, it would mean that the core difficulty of continual learning in this domain is largely handled by the pretrained representation, and that future work should start from foundation models rather than from scratch.

What carries the argument

The mechanism is the frozen foundation model as a fixed feature extractor: images are mapped to pre-trained embeddings $z = F_\theta(x)$ that are never fine-tuned. On top, per-task lightweight MLP heads are trained independently and concatenated at inference, or class-mean prototypes $\mu_c$ are stored in a memory bank and matched by nearest-neighbor distance. The paper also uses data transformations on the embedding space, $\ell^2$ normalization and a learnable hyperbolic projection, to improve the prototype classifier.

What would settle it

Compare the pretraining image sets of Google Derm and PanDerm against the three evaluation datasets; if any evaluation image or near-duplicate appears in pretraining, the claimed advantage over methods that never saw those images is inflated. A cleaner check is to rerun the identical CIL protocol on a skin-lesion dataset released after the models' pretraining cutoff and see whether the frozen-MLP result still stands.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the continual-learning problem in dermatology largely disappears when the feature extractor is a frozen foundation model trained on large-scale skin imagery. With PanDerm embeddings and a two-hidden-layer MLP head trained only on current-task data, the model reaches a balanced accuracy of 92.25% on HAM10000, surpassing the SINGLE reference model (88.35%), and reports zero forgetting on all three datasets; the same pattern holds on Dermofit (93.11%) and Derm7pt (77.80%). A prototype-based nearest-mean classifier with no trained parameters also produces competitive results on some benchmarks, and a learnable hyperbolic projection boosts it substantially (e.g., 64.75% to 81.41% on HAM with Google Derm). The authors interpret these results as evidence that rich frozen features, rather than specialized continual-learning machinery, are what drive stability in class-incremental settings.

Load-bearing premise

The foundation models were pretrained on massive dermatology image collections that could include the exact HAM10000, Dermofit, and Derm7pt images used for evaluation, and the paper does not rule out that overlap.

Editorial extensions

If this is right

On the three benchmarks tested, training a small MLP on frozen embeddings eliminates catastrophic forgetting while outperforming every compared continual-learning method, so replay and regularization appear unnecessary for this setting.
A general-purpose CLIP ViT-L/14 backbone with an MLP head also beats all previous continual-learning methods on these datasets, suggesting the benefit is not unique to dermatology-specific pretraining.
Prototype-based zero-training classifiers depend heavily on the alignment between pretraining and target domain: they lag with CLIP but improve markedly with dermatology-specific embeddings and embedding-space projections.
The paper reports that the frozen-MLP approach exceeds the SINGLE upper bound on HAM (92.25% vs 88.35%), which would mean incremental training on frozen features costs nothing relative to task-specialized models.

Reading between the lines

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

Inference: If frozen-embedding classifiers truly achieve zero forgetting and beat the single-task upper bound, then a large share of continual-learning research for medical imaging may be attacking a problem that pretrained representations already solve; the remaining open question is representation freshness as new disease classes appear over time.
Inference: Because the method stores only class-mean prototypes or lightweight MLP weights rather than raw patient images, it offers a natural fit for clinical privacy constraints, though the same property means the memory bank cannot be corrected if the embedding space is biased.
Inference: A decisive test the paper leaves implicit is whether its evaluation datasets overlap with the foundation models' pretraining data; running the same protocol on a dermatology dataset released after the models' training cutoff would settle whether the advantage is generalization or memorization.

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.

Desk Editor's Note

read the letter

Short version: this is a clean, systematic baseline comparison for CIL in dermatology, and the central caution is justified — the derm-specific FMs may have seen the test data. If that overlap is real, the Table 1 results are not valid evidence.

What's genuinely useful: the authors evaluate two frozen derm FMs and CLIP on three public skin lesion sets, under a fixed CIL protocol taken from Continual-Zoo. The main MLP baseline (frozen features + per-task MLP) is simple and beats the replay/regularization competitors. The NMC ablation, including normalization, random projection, hyperbolic projection, and LDA, is reasonably thorough. That part gives a practical map of what works with prototype classifiers on these embeddings.

The soft spots are in proportion. The biggest is the leakage question. Google Derm and PanDerm are trained on large derm image collections; HAM, Dermofit, and Derm7pt are standard public sets. The paper gives no overlap analysis, no membership test, no discussion. Without that, 'PanDerm MLP beats the SINGLE upper bound on HAM' is not an out-of-distribution result — it could simply be the FM memorizing test images. That is load-bearing. The CLIP results are a useful counterpoint, since CLIP is general-purpose and less likely to have seen these exact images, but the paper does not frame the main claim around CLIP.

The other issues matter but are smaller. The code link is missing. MLP hidden sizes, task splits, and the hyperbolic projection settings are not fully specified. The comparison against prior methods relies entirely on numbers from the authors' own Continual-Zoo paper, so a skeptical reader can't verify those baselines independently.

Bottom line: the paper deserves a serious referee, because the question is real and the experimental design is mostly sound. But it should be accepted conditionally, with a required leakage analysis and release of code and exact splits. As it stands, I would not cite the derm-FM numbers without first checking overlap.

Referee Report

4 major / 5 minor

Summary. The paper evaluates frozen dermatology foundation models (Google Derm and PanDerm) as feature extractors for class-incremental learning on three public skin-lesion datasets: HAM10000, Dermofit, and Derm7pt. It proposes two baselines: a per-task MLP classifier trained on frozen embeddings, and a prototype-based nearest-mean classifier that requires no training. The authors report that the MLP baseline achieves balanced accuracy of 92.25% on HAM10000 with zero forgetting, outperforming all compared continual learning methods and also the SINGLE upper-bound reference. Ablation studies explore NMC variants (normalization, random projection, hyperbolic projection, PCA, LDA) and replace the dermatology-specific models with CLIP ViT-L/14. The paper concludes that frozen foundation models should be the default starting point for continual learning in dermatology.

Significance. If the empirical claims hold, the paper would provide a practically important result for medical continual learning: a frozen foundation model with a lightweight per-task classifier can outperform specialized regularization, replay, and architecture-based methods while avoiding the privacy and storage issues of replay. The paper is among the first to evaluate dermatology-specific foundation models in a class-incremental setting, and it includes ablations across multiple feature extractors and classifier variants. The central claims, however, rest entirely on the validity of the evaluation protocol, and the manuscript currently does not rule out pretraining overlap between the foundation models and the evaluation datasets, nor does it provide enough protocol detail to verify the comparisons against prior work.

major comments (4)

[Section 4.1, Implementation Details] The pretraining corpora of Google Derm and PanDerm are not analyzed for overlap with HAM10000, Dermofit, or Derm7pt. PanDerm is described as pretrained on millions of clinical and dermoscopic dermatology images, and these three benchmarks are standard public skin-lesion collections; if any evaluation images or near-duplicates appeared in pretraining, the Table 1 results are inflated and the comparison with methods that did not use those features is unfair. This is load-bearing because the paper's headline result of surpassing the SINGLE upper bound depends on the evaluation being out-of-distribution for the frozen encoders. Please provide an overlap analysis, cite pretraining data documentation, or use models with explicit exclusion guarantees.
[Section 4.1, Reference Methods and Competitors; Table 1] All non-FM baseline numbers in Table 1 are taken from Continual-Zoo [8], but the manuscript does not state whether the train/test splits, task order, image preprocessing, and evaluation code are exactly those used in [8] or were re-implemented. If any protocol detail differs, the reported gains over Continual-Zoo may reflect protocol differences rather than the frozen-FM effect. Please provide the exact task partitions, the number of tasks T, and the version of [8]'s code used; ideally, re-run the baselines with the same evaluation harness. Also clarify the backbone used for the SINGLE upper bound, since a SINGLE model trained on a weaker backbone is not an upper bound for a method that uses a stronger frozen feature extractor.
[Section 3.2, Inference Phase; Table 1] The zero forgetting (F=0) reported for the MLP baseline is guaranteed by construction, because each task receives a newly initialized head and old heads are never updated or regularized; the same holds for the NMC prototypes. Reporting F=0 as an empirical finding therefore overstates the result. The meaningful comparison is the final balanced accuracy under the task-agnostic inference rule (arg max over concatenated heads), and this should be stated explicitly in the main text.
[Section 4.1, Implementation Details; abstract] The MLP hidden layer sizes, the number of tasks T, the class partition for each dataset, and the validation split are not reported, and the promised code link is absent. These details are necessary to reproduce the central results, especially because the method's advantage is empirical and protocol-dependent. Please specify these items or provide a public repository with the exact configuration.

minor comments (5)

[Section 4.2, Ablation 1] The hyperbolic projection is described as having a significant positive impact, but on DMF it decreases accuracy relative to the base NMC for both models (Google Derm 63.79 vs 67.56; PanDerm 43.21 vs 49.27); the sentence should be qualified to specific datasets.
[Section 4.1, Implementation Details] The sentence 'Our code and datasets are available here' contains no link or reference; please add a URL or footnote.
[References] Reference [10] appears unrelated to continual learning (it is a survey of glucose monitoring systems); please verify the citation and replace it with the intended work.
[Table 2 and Table 3] The column header 'Derm' should be written as 'Google Derm' for consistency with the main text.
[Section 4.2, Results and Analysis] The statement that NMC-based baselines 'achieve comparable, and sometimes superior, results' is too strong: on DMF and D7P with PanDerm, the base NMC achieves 49.27% and 44.51%, respectively, which is below most competing methods; please calibrate this sentence.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central result is an empirical benchmark comparison, and the only self-referential element is a non-load-bearing adoption of the Continual-Zoo experimental protocol.

full rationale

The paper's central claim is an empirical comparison: frozen dermatology foundation models with a lightweight MLP or prototype classifier outperform existing continual learning methods on three public benchmarks. This claim is established through external baselines and measured accuracy, not through a derivation that reduces to its own inputs. The only self-referential element is the adoption of dataset splits and protocol from Continual-Zoo [8], whose author list overlaps with the present paper; however, that protocol is an independent experimental setting, and the reported accuracies do not reduce to it or depend on it for their validity. The zero-forgetting values (F = 0) are a direct consequence of the frozen-backbone and independent-head design, but the paper presents them as measured metrics rather than as a prediction derived from a fitted parameter. The learnable hyperbolic projection is an ablation component with parameters optimized on training data, not a fitted input renamed as a prediction. The concern that PanDerm or Google Derm may have been pretrained on evaluation images is a data-leakage validity risk, not a logical circularity, and would affect correctness rather than the derivation chain. No equation, definition, or self-citation chain makes the central result true by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that the FMs did not see the test data and that the comparison to prior literature is fair. The MLP baseline introduces standard hyperparameters but they are under-specified. The hyperbolic projection is an ablation component with fitted parameters, not central to the main result.

free parameters (3)

MLP hidden layer sizes = not specified
The paper does not give the number of units in the two hidden layers of the MLP head. This is a hand-chosen architectural parameter that affects the results.
MLP optimization hyperparameters = learning rate 0.001, batch size 200, up to 200 epochs with early stopping
These hyperparameters are chosen by the authors and may affect performance. They are standard for this type of model, but are not justified or swept.
Hyperbolic projection parameters = optimized during training
The learnable hyperbolic projection from [17] introduces fitted parameters. It is used in the NMC ablation studies, not in the central MLP baseline, but the paper claims it improves results.

assumptions (4)

domain assumption The three public datasets are representative of the clinical deployment distribution and the task partitioning is meaningful for CIL.
Section 4.1 uses HAM10000, Dermofit, and Derm7pt as benchmarks without discussing distribution shift or clinical relevance beyond the class split.
domain assumption The Google Derm and PanDerm foundation models were not pretrained on the evaluation images.
Section 4.1 introduces these models as pretrained on large dermatology collections but does not confirm they exclude HAM10000, Dermofit, and Derm7pt. The correctness of the main result depends on this.
domain assumption The numbers reported for prior methods from Continual-Zoo [8] are comparable to the authors' numbers.
The paper says it adopts the dataset splits and protocol from [8], but does not restate them or verify that all competing methods used the same feature extractors and task orders.
domain assumption Balanced accuracy (BAAC) is the appropriate evaluation metric.
Section 4.1 defines BAAC as the primary metric. The paper does not discuss whether other metrics would change the conclusions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Foundation Models as Class-Incremental Learners for Dermatological Image Classification." pith.science (2026). https://pith.science/paper/ONP4CGGN

@misc{pith2026250714050,
  author       = {Pith},
  title        = {Pith review of: Foundation Models as Class-Incremental Learners for Dermatological Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ONP4CGGN}},
  note         = {Machine review of arXiv:2507.14050}
}

read the original abstract

Class-Incremental Learning (CIL) aims to learn new classes over time without forgetting previously acquired knowledge. The emergence of foundation models (FM) pretrained on large datasets presents new opportunities for CIL by offering rich, transferable representations. However, their potential for enabling incremental learning in dermatology remains largely unexplored. In this paper, we systematically evaluate frozen FMs pretrained on large-scale skin lesion datasets for CIL in dermatological disease classification. We propose a simple yet effective approach where the backbone remains frozen, and a lightweight MLP is trained incrementally for each task. This setup achieves state-of-the-art performance without forgetting, outperforming regularization, replay, and architecture based methods. To further explore the capabilities of frozen FMs, we examine zero training scenarios using nearest mean classifiers with prototypes derived from their embeddings. Through extensive ablation studies, we demonstrate that this prototype based variant can also achieve competitive results. Our findings highlight the strength of frozen FMs for continual learning in dermatology and support their broader adoption in real world medical applications. Our code and datasets are available here.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

LesionGen: A Concept-Guided Diffusion Model for Dermatology Image Synthesis
eess.IV 2025-07 conditional novelty 5.0 of 10

Concept-guided captions and prompt balancing improve synthetic skin lesion images enough that augmenting real data with them lifts rare-class classification, though synthetic-only training remains clearly worse than r...

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages · cited by 1 Pith paper

[8]

In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

Bayasi, N., Hamarneh, G., Garbi, R.: Continual-zoo: Leveraging zoo models for continual classification of medical images. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 4128–4138 (2024)

2024
[1]

Medical Image Analysis (2024)

Ayromlou, S., Tsang, T., Abolmaesumi, P., Li, X.: CCSI: Continual class-specific impression for data-free class incremental learning. Medical Image Analysis (2024)

work page 2024
[2]

In: Color Medical Image Analysis, pp

Ballerini, L., Fisher, R.B., Aldridge, B., Rees, J.: A color and texture based hierar- chical k-nn approach to the classification of non-melanoma skin lesions. In: Color Medical Image Analysis, pp. 63–86. Springer (2013)

work page 2013
[3]

In: IEEE/CVF CVPR

Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learn- ing with a memory of diverse samples. In: IEEE/CVF CVPR. pp. 8218–8227 (2021)

work page 2021
[4]

Bayasi, N.: Beyond catastrophic forgetting: advancing continual learning for robust andfairmedicalimageanalysis.Ph.D.thesis,UniversityofBritishColumbia(2025)

work page 2025
[5]

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

Bayasi, N., Du, S., Hamarneh, G., Garbi, R.: Continual-gen: Continual group en- sembling for domain-agnostic skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 3–13 (2023)

2023
[6]

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

Bayasi, N., Fayyad, J., Bissoto, A., Hamarneh, G., Garbi, R.: Biaspruner: Debiased continual learning for medical image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 90–101. Springer (2024)

2024
[7]

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

Bayasi, N., Hamarneh, G., Garbi, R.: Culprit-prune-net: Efficient continual se- quential multi-domain learning with application to skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 165–175. Springer (2021)

2021

Show all 36 references

[9]

IEEE Transactions on Medical Imaging (2024)

Bayasi, N., Hamarneh, G., Garbi, R.: GC2: Generalizable continual classification of medical images. IEEE Transactions on Medical Imaging (2024)

2024
[10]

In: 2013 IEEE 20th International Confer- ence on Electronics, Circuits, and Systems (ICECS)

Bayasi, N., Saleh, H., Mohammad, B., Ismail, M.: The revolution of glucose mon- itoring methods and systems: A survey. In: 2013 IEEE 20th International Confer- ence on Electronics, Circuits, and Systems (ICECS). pp. 92–93 (2013)

2013
[11]

In: (MICCAI)

Bera, S., Ummadi, V., Sen, D., Mandal, S., Biswas, P.K.: Memory replay for contin- ual medical image segmentation through atypical sample selection. In: (MICCAI). pp. 513–522 (2023)

2023
[12]

arXiv preprint arXiv:2108.07258 (2021)

Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

2021 arXiv
[13]

In: Proceedings of the AAAI conference on artificial intelligence

Chee, E., Lee, M.L., Hsu, W.: Leveraging old knowledge to continually learn new classes in medical images. In: Proceedings of the AAAI conference on artificial intelligence. vol. 37, pp. 14178–14186 (2023)

2023
[14]

arXiv preprint arXiv:2405.16328 (2024)

Chen, X., Zheng, H., Xie, Y., Ma, Y., Li, T.: A classifier-free incremental learning framework for scalable medical image segmentation. arXiv preprint arXiv:2405.16328 (2024)

2024 arXiv
[15]

Fayyad, J.: Out-of-distribution detection using inter-level features of deep neural networks. Ph.D. thesis, University of British Columbia (2023)

2023
[16]

Computer Methods and Programs in Biomedicine 253, 108231 (2024)

Fayyad, J., Alijani, S., Najjaran, H.: Empirical validation of conformal prediction for trustworthy skin lesions classification. Computer Methods and Programs in Biomedicine 253, 108231 (2024)

2024
[17]

Gonzalez-Jimenez, A., Lionetti, S., Amruthalingam, L., Gottfrois, P., Gröger, F., Pouly, M., Navarini, A.A.: Is hyperbolic space all you need for medical anomaly detection? (2025), provisionally accepted at MICCAI 2025 10 Elkhayat & Mahmoud et al

2025
[18]

https://developers.google.com/ health-ai-developer-foundations/derm-foundation (2025)

Google Health AI: Derm foundation model. https://developers.google.com/ health-ai-developer-foundations/derm-foundation (2025)

2025
[19]

In: NeurIPS (2023)

Janson, P., Zhang, W., Aljundi, R., Elhoseiny, M.: A simple baseline that questions the use of pretrained models in continual learning. In: NeurIPS (2023)

2023
[20]

In: CVPR (2023)

Javed, K., Wang, Y., Xu, Q., Zhang, Y.: Parametric prompt tuning for vision- language models. In: CVPR (2023)

2023
[21]

The IEEE Journal of Biomedical and Health Informatics23(2), 538–546 (2018)

Kawahara, J., Daneshvar, S., Argenziano, G., Hamarneh, G.: Seven-point checklist and skin lesion classification using multitask multimodal neural nets. The IEEE Journal of Biomedical and Health Informatics23(2), 538–546 (2018)

2018
[22]

Proceedings of the National Academy of Sciences 114(13), 3521–3526 (2017)

Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 114(13), 3521–3526 (2017)

2017
[23]

IEEE Transactions on Pattern Analysis and Machine Intelligence40(12), 2935–2947 (2017)

Li, Z., Hoiem, D.: Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence40(12), 2935–2947 (2017)

2017
[24]

arXiv preprintarXiv:2307.02251(2023)

McDonnell, M.D., Gong, D., Parveneh, A., Abbasnejad, E., van den Hengel, A.: Ranpac: Random projections and pre-trained models for continual learning. arXiv preprintarXiv:2307.02251(2023). https://doi.org/10.48550/arXiv.2307.02251

2023 doi arXiv
[25]

In: CVPR

Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remem- ber: A synaptic plasticity driven framework for continual learning. In: CVPR. pp. 11321–11329 (2019)

2019
[26]

Psychological Review97(2), 285 (1990)

Ratcliff, R.: Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological Review97(2), 285 (1990)

1990
[27]

In: IEEE/CVF CVPR

Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: iCaRL: Incremental clas- sifier and representation learning. In: IEEE/CVF CVPR. pp. 2001–2010 (2017)

2017
[28]

Scientific data 5(1), 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 5(1), 1–9 (2018)

2018
[29]

Nature communications11(1), 1–14 (2020)

van de Ven, G.M., Siegelmann, H.T., Tolias, A.S.: Brain-inspired replay for con- tinual learning with artificial neural networks. Nature communications11(1), 1–14 (2020)

2020
[30]

In: ECCV

Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., et al.: Dualprompt: Complementary prompting for rehearsal-free continual learning. In: ECCV. pp. 631–648 (2022)

2022
[31]

In: IEEE/CVF CVPR

Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T.: Learning to prompt for continual learning. In: IEEE/CVF CVPR. pp. 139–149 (2022)

2022
[32]

Signal Processing (2024)

Wu, Z., Zhu, F., Guo, K., Sheng, R., Chao, L., Fang, H.: Modal adaptive super- resolution for medical images via continual learning. Signal Processing (2024)

2024
[33]

arXiv preprint arXiv:2410.15038 (2024)

Yan, S., Yu, Z., Primiero, C., Vico-Alonso, C., Wang, Z., Yang, L., Tschandl, P., Hu, M., Tan, G., Tang, V., Ng, A.B., Powell, D., Bonnington, P., See, S., Janda, M., Mar, V., Kittler, H., Soyer, H.P., Ge, Z.: A general-purpose multimodal foundation model for dermatology. arXi...

2024 arXiv
[34]

Visual Intelli- gence 1(1), 5 (2023)

Yang, Y., Cui, Z., Xu, J., Zhong, C., Zheng, W.S., Wang, R.: Continual learning with bayesian model based on a fixed pre-trained feature extractor. Visual Intelli- gence 1(1), 5 (2023)

2023
[35]

In: MICCAI

Zhang, W., Huang, Y., Zhang, T., Zou, Q., Zheng, W.S., Wang, R.: Adapter learn- ing in pretrained feature extractor for continual learning of diseases. In: MICCAI. pp. 68–78 (2023)

2023
[36]

In: MICCAI

Zhang, Y., Li, X., Chen, H., Yuille, A.L., Liu, Y., Zhou, Z.: Continual learning for abdominal multi-organ and tumor segmentation. In: MICCAI

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [8]

In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

Bayasi, N., Hamarneh, G., Garbi, R.: Continual-zoo: Leveraging zoo models for continual classification of medical images. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 4128–4138 (2024)

2024

[2] [1]

Medical Image Analysis (2024)

Ayromlou, S., Tsang, T., Abolmaesumi, P., Li, X.: CCSI: Continual class-specific impression for data-free class incremental learning. Medical Image Analysis (2024)

work page 2024

[3] [2]

In: Color Medical Image Analysis, pp

Ballerini, L., Fisher, R.B., Aldridge, B., Rees, J.: A color and texture based hierar- chical k-nn approach to the classification of non-melanoma skin lesions. In: Color Medical Image Analysis, pp. 63–86. Springer (2013)

work page 2013

[4] [3]

In: IEEE/CVF CVPR

Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learn- ing with a memory of diverse samples. In: IEEE/CVF CVPR. pp. 8218–8227 (2021)

work page 2021

[5] [4]

Bayasi, N.: Beyond catastrophic forgetting: advancing continual learning for robust andfairmedicalimageanalysis.Ph.D.thesis,UniversityofBritishColumbia(2025)

work page 2025

[6] [5]

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

Bayasi, N., Du, S., Hamarneh, G., Garbi, R.: Continual-gen: Continual group en- sembling for domain-agnostic skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 3–13 (2023)

2023

[7] [6]

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

Bayasi, N., Fayyad, J., Bissoto, A., Hamarneh, G., Garbi, R.: Biaspruner: Debiased continual learning for medical image classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 90–101. Springer (2024)

2024

[8] [7]

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

Bayasi, N., Hamarneh, G., Garbi, R.: Culprit-prune-net: Efficient continual se- quential multi-domain learning with application to skin lesion classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 165–175. Springer (2021)

2021