A Survey on Deep Learning of Small Sample in Biomedical Image Analysis

Pengyi Zhang; Xiaoqiong Li; Xiaoying Tang; Yulin Deng; Yunxin Zhong

REVIEW 3 major objections 5 minor 2 cited by

A Survey on Deep Learning of Small Sample in Biomedical Image Analysis

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Small-sample deep learning in biomedical imaging is best attacked through five complementary technique families.

desk verdict Useful survey with an unsupported 'comprehensive' claim; needs revision, not rejection. read the letter →

arxiv 1908.00473 v1 pith:RCIDWEN3 submitted 2019-08-01 cs.CV

Pengyi Zhang , Yunxin Zhong , Yulin Deng , Xiaoying Tang , Xiaoqiong Li This is my paper · ORCID

classification cs.CV

keywords smallsamplelearningbiomedicalimageanalysisdeepexplanationmethodsweaklysupervisedtransferactivedataaugmentation

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

Small-sample learning is a practical barrier in biomedical image analysis because expert annotations are scarce, expensive, and task-specific. This survey argues that the barrier can be lowered by five families of techniques: explanation, weakly supervised learning, transfer learning, active learning, and miscellaneous data-side methods such as augmentation, domain knowledge, shallow-method fusion, and attention. Its central claim is that these techniques, mostly developed in computer vision, can be adapted to clinical biomedical imaging and together keep deep models usable when large labeled sets are unavailable. The paper positions itself as the first structured survey of key small-sample techniques written specifically for deep learning in clinical biomedical image analysis, complementing earlier reviews that mention the small-sample problem only as a challenge.

What carries the argument

The organizing mechanism is a five-category taxonomy tied to a deep-learning workflow. Explanation targets the clinical-decision stage, weakly supervised learning targets annotation cost, transfer learning targets the training stage, active learning targets dataset collection, and the miscellaneous group targets the model and data themselves. Inside the categories, recurring technical devices carry the argument: class activation maps and gradient-based saliency for explanation and weak localization; an iterative pseudo-label loop of initial labels, network training, and post-processing refinement for weak supervision; fine-tuning, multi-task sharing, and adversarial domain adaptation for transfer; uncertainty, diversity, and committee-based querying for active learning; and transformations, synthetic images from generative models, and learned augmentation policies for data expansion.

What would settle it

Finding a substantial class of small-sample techniques used in biomedical imaging, such as self-supervised pre-training on unlabeled scans, that fits none of the five categories, or reproducing a cited study and finding its gain over training from scratch disappears, would undermine the survey's central claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that the small-sample learning dilemma in biomedical image analysis is not one problem but several, each with a workable remedy. A practitioner facing scarce annotations should choose among five technique families: explanation methods that make decisions transparent and can double as localizers; weakly supervised learning that turns cheap coarse annotations into fine-grained predictions; transfer learning that reuses knowledge from data-rich domains; active learning that spends the annotation budget on the most informative samples; and miscellaneous techniques: data augmentation, domain knowledge, shallow-method fusion, and attention, which regularize or enlarge the effective training set. The survey states that this is the first attempt to bring these key SSL techniques together for clinical biomedical image analysis, and supports each category with representative studies across fundus photos, chest X-rays, OCT, histology, MRI, CT, and ultrasound.

Load-bearing premise

The survey's usefulness rests on its five-category taxonomy actually covering the techniques that matter for small-sample biomedical deep learning, and on the cited papers' reported gains being accurate and reproducible.

Editorial extensions

If this is right

Practitioners with scarce annotations can expect at least one of the five routes to fit their task, rather than treating small-sample learning as an unsolved blocker.
Cheap annotation modalities such as image-level labels, boxes, scribbles, and points become a viable path to pixel-level segmentation and lesion localization.
Fine-tuning a network pre-trained on a large natural-image dataset should generally beat training from scratch on biomedical tasks.
Active learning should reach a target performance with fewer labels than random sampling, especially when combined with transfer learning.
Data augmentation, including learned augmentation, is a primary lever for small-sample segmentation and classification.

Reading between the lines

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

The paper presents the five categories as separate, but a natural extension is that they compound: explanations can generate weak labels, active learning can feed transfer learning, and augmentation can amplify both. A benchmark that stacks the categories against any single one would test whether the combination is the real win.
Because the survey predates the recent wave of self-supervised pre-training on unlabeled medical images, an updated taxonomy would probably add that as a sixth route; if such methods keep improving, the five-category claim will age.
The explanation section implies a testable hypothesis: visual explanations that agree with clinician-identified regions should correspond to more trustworthy models. Measuring whether class-activation-map agreement with radiologist annotations predicts diagnostic accuracy would check that.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Dear [Colleague],

This survey is a useful annotated map of selected small-sample learning techniques for biomedical image analysis, but its central selling point—being the first comprehensive survey—does not survive contact with its own table of contents. Read it as a curated overview of five technique families, not as a complete treatment of the field.

What the paper does well: it is clearly organized around five categories (explanation, weakly supervised learning, transfer learning, active learning, and a miscellaneous bucket of augmentation, domain knowledge, shallow methods, attention). The tables give a quick entry point into representative papers per category, with datasets and model details. The figures, especially the overview diagram, are helpful for a newcomer. The authors also point to a GitHub repository with demos, which is a nice touch. For a reader who wants to know 'what kinds of approaches exist for training deep models on small biomedical datasets,' this delivers a workable starting point.

The soft spots are real and proportionate to the claims. First, the 'first comprehensive survey' assertion in Sections 1 and 7 is not supported. The taxonomy omits meta-learning and few-shot learning entirely—arguably the most direct small-sample learning techniques, already prominent by 2019 (MAML, prototypical networks). The only 'one-shot' mention is one paper under learned augmentation. This is not a pedantic point: the paper's own differentiator is comprehensiveness. Second, the inclusion of explanation methods is acknowledged as not SSL, which is fine as a clinical-practice extension, but it blurs the scope. Third, there are citation inconsistencies: duplicate Rajpurkar 2018 entries, a mismatch between Zhou et al. 2018 and 2019, and minor formatting errors. These are not fatal, but they undermine reliability as a reference.

None of this kills the paper's value as a structured overview. It is a legitimate entry point for clinicians or researchers new to the area, and the authors have clearly engaged with a broad literature. But the claims need to be recalibrated: either add the missing technique families or revise the 'first comprehensive' language to something like 'a structured survey of key techniques selected for clinical relevance.'

For peer review: I would send it out rather than desk-reject, because it fills a niche and the flaws are fixable with a substantive revision. The taxonomy and scope decisions need discussion, and the citation errors need cleaning. A good referee could help the authors turn this into a genuinely useful reference.

Referee Report

3 major / 5 minor

Summary. This paper is a survey of techniques intended to support deep learning for biomedical image analysis when labeled samples are scarce, which the authors term the 'small sample learning' (SSL) dilemma. The survey organizes key techniques into five categories: explanation methods, weakly supervised learning, transfer learning, active learning, and a miscellaneous group covering data augmentation, domain knowledge, shallow methods, and attention mechanisms. Each category is described with representative papers, many summarized in tabular form, and the paper includes a GitHub repository with demos. The central claim, stated in Sections 1 and 7, is that this is the first comprehensive survey of SSL techniques for clinical biomedical image analysis.

Significance. The paper's value would lie in providing a structured entry point to a broad set of techniques for small-sample biomedical image analysis, with useful tables that map methods to tasks, datasets, and network architectures. The organization around five categories is readable and the included demos are a practical addition. However, the significance rests on the claim of comprehensiveness, and that claim is currently not supported because the survey omits entire major branches of small-sample learning, notably meta-learning/few-shot learning and self-supervised learning, while including a category (explanation techniques) that the authors themselves state is not an SSL method. If the scope is reframed and the missing branches are addressed, the survey could be a useful contribution; as written, it is better described as an annotated overview of selected techniques.

major comments (3)

[Sections 1 and 7; Fig. 1; Table 1] The claim that this is the 'first comprehensive survey on key SSL techniques' is not supported as written. The five-category taxonomy (explanation, weakly supervised learning, transfer learning, active learning, and miscellaneous) omits meta-learning/few-shot learning and self-supervised learning, which by 2019 were established and widely used approaches for small-sample image analysis (e.g., MAML, prototypical networks, contrastive learning). The paper itself cites Shu et al. (2018), whose SSL framework includes 'experience learning,' under which meta-learning naturally falls; omitting this branch undermines the stated novelty, which rests on comprehensiveness. The authors should either add a section covering few-shot/meta-learning and self-supervised learning or revise the claim to describe the survey as covering selected key techniques rather than being comprehensive. This is a load-bearing issue because the same claim appears in both the introduction and the conclusion.
[Abstract, Section 2, and Section 7] Explanation techniques are not SSL techniques, and the paper acknowledges this in the abstract ('we intentionally expand this survey to include the explanation methods') and in the introduction. Including them in a survey of SSL techniques blurs the scope and weakens the taxonomy. The authors should either justify more explicitly why explanation methods belong in an SSL survey, or present them as a separate clinical-facing extension rather than one of the five categories of SSL techniques. This affects the paper's internal consistency and the accuracy of its stated categorization in Fig. 1 and the abstract.
[Tables 1, 3, and 4; Sections 4 and 5] There are multiple citation inconsistencies that affect the reliability of the tables. In Table 1, Rajpurkar et al. (2018) appears twice with different descriptions; based on the text and references, one of these should likely be Rajpurkar et al. (2017) (the MURA musculoskeletal radiograph paper) and the other Rajpurkar et al. (2018) (the CheXNeXt chest X-ray paper). In Table 3, a 'Zhou et al. (2018)' entry for carotid intima-media thickness video interpretation is cited, but the reference list contains only Zhou et al. (2017) and Zhou et al. (2019), and Section 4 consistently cites the latter two. Similarly, in Table 4, 'Zhou et al. (2018)' appears for the same carotid video interpretation task, while the text and references point to Zhou et al. (2019). Table 5 also contains the ambiguous entry 'Xie et al. (2018)(2019)'. These errors must be corrected and a single consistent citation style used throughout the tables and text.

minor comments (5)

[General] There are several typos and awkward phrasings, for example 'We bulid demos' in the abstract and 'furtherly improve' in multiple places. A careful proofread is recommended.
[Section 2 and Table 1] The sentence 'Rajpurkar et al. (2018)(2017) adopted this approach...' is confusing; it should list the two citations separately with the appropriate years and tasks.
[Section 3 and Figure 4] The text states that the gap between weakly supervised object detection and YOLO is 'less than 10%', but the figure shows a gap of 0.094 (9.4%). This is correct in magnitude, but the caption should clarify that this is an absolute mAP difference and that the comparison is to a single, non-ensemble YOLO model.
[Section 6.3 and Table 7] Table 7 lists 'Schlemper et al. (2018)' twice with different methods (attention-gated branches for ultrasound plane detection and attention-gated skip connections for pancreas segmentation). These are indeed two related papers commonly cited together, but the duplicate author-year entries may confuse readers; disambiguating them with a, b or additional author initials would improve clarity.
[References] The reference list contains inconsistencies between the citation years used in the text and the reference years, such as the Zhou et al. (2018)/(2019) issue noted above. The references should be checked to ensure every in-text citation has a matching reference entry and vice versa.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the survey's claims concern coverage and categorization, not a derivation from its own inputs.

full rationale

This manuscript is a literature survey rather than a derivation or empirical study, so there is no chain of equations, fitted parameters, or imported uniqueness theorems whose outputs could reduce to their inputs. The central novelty claim ('To our best knowledge, this is the first attempt to present a comprehensive survey') is an assertion about coverage; it can be challenged on completeness grounds, for instance because meta-learning and few-shot learning are not given a dedicated category, but that is a correctness or scope concern, not circular reasoning. The claim is not used as evidence for itself, and no load-bearing step is justified by a self-citation. The cited prior reviews and technique papers are external to the authors and are used descriptively. The inclusion of explanation methods is explicitly an editorial expansion rather than a consequence of the small-sample-learning definition, and the paper acknowledges that these methods are not SSL techniques proper. Under the rule that only demonstrable reductions count as circularity, the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

No free parameters or invented entities are present. The survey's content is purely descriptive and rests on the cited literature.

assumptions (1)

domain assumption Deep learning models for biomedical image analysis suffer from a shortage of labeled data, motivating small-sample techniques.
Stated in the abstract and introduction as the starting premise of the survey.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Survey on Deep Learning of Small Sample in Biomedical Image Analysis." pith.science (2026). https://pith.science/paper/RCIDWEN3

@misc{pith2026190800473,
  author       = {Pith},
  title        = {Pith review of: A Survey on Deep Learning of Small Sample in Biomedical Image Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RCIDWEN3}},
  note         = {Machine review of arXiv:1908.00473}
}

read the original abstract

The success of deep learning has been witnessed as a promising technique for computer-aided biomedical image analysis, due to end-to-end learning framework and availability of large-scale labelled samples. However, in many cases of biomedical image analysis, deep learning techniques suffer from the small sample learning (SSL) dilemma caused mainly by lack of annotations. To be more practical for biomedical image analysis, in this paper we survey the key SSL techniques that help relieve the suffering of deep learning by combining with the development of related techniques in computer vision applications. In order to accelerate the clinical usage of biomedical image analysis based on deep learning techniques, we intentionally expand this survey to include the explanation methods for deep models that are important to clinical decision making. We survey the key SSL techniques by dividing them into five categories: (1) explanation techniques, (2) weakly supervised learning techniques, (3) transfer learning techniques, (4) active learning techniques, and (5) miscellaneous techniques involving data augmentation, domain knowledge, traditional shallow methods and attention mechanism. These key techniques are expected to effectively support the application of deep learning in clinical biomedical image analysis, and furtherly improve the analysis performance, especially when large-scale annotated samples are not available. We bulid demos at https://github.com/PengyiZhang/MIADeepSSL.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

Beyond Random Partitioning: Unsupervised Spatio-Temporal Stratification for Cohort Balancing in Longitudinal Medical Imaging
cs.CV 2026-07 conditional novelty 4.0 of 10

K-means clustering on six intensity and temporal features plus intra-cluster stratified sampling reduces cross-subset imaging and temporal imbalance versus random splitting in a 149-patient longitudinal brain MRI cohort.
Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation
eess.IV 2019-08 conditional novelty 3.0 of 10

A structured review of deep learning segmentation techniques for scarce and weak annotations, with cost-gain recommendations.

Reference graph

Works this paper leans on

5 extracted references · 1 canonical work pages · cited by 2 Pith papers

[53]

https://doi.org/10.1093/nsr/nwx106

doi:10.1093/nsr/nwx106
[191]

Why Should I Trust You?

https://doi.org/10.1007/978-3-030-00889-5_21 Pan, Y ., Liu, M., Lian, C., Zhou, T., Xia, Y ., Shen, D., 2018. Synthesizing Missing PET from MRI with Cycle-consistent Generative Adversarial Networks for Alzheimer ’s Disease Diagnosis, in: Frangi, A.F., Schnabel, J.A., Davatzikos, C., Alberola-López, C., Fichtinger, G. (Eds.), Medical Image Computing and Co...

arXiv 2018
[2600]

Unsupervised Domain Adaptation in Brain Lesion Segmentation with Adversarial Networks, in: Niethammer, M., Styner, M., Aylward, S., Zhu, H., Oguz, I., Yap, P

https://doi.org/10.1109/TCSVT.2016.2589879 Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V ., Simpson, J., Kane, A., Menon, D., Nori, A., Criminisi, A., Rueckert, D., Glocker, B., 2017a. Unsupervised Domain Adaptation in Brain Lesion Segmentation with Adversarial Networks, in: Niethammer, M., Styner, M., Aylward, S., Zhu, H., Oguz, I., Yap, P. -T.,...

arXiv 2016
[2854]

Nutrition and Care of the Preterm Infant: Current Issues

https://doi.org/10.1109/CVPR.2016.311 Bloch, N., Madabhushi, A., Huisman, H., 2015. NCI-ISBI 2013 Challenge: Automated Segmentation of Prostate Structures. The Cancer Imaging Archive 370. https://doi.org/10.7937/k9/tcia.2015.zf0vlopv Bonechi, S., Andreini, P., Bianchini, M., Scarselli, F., 2018. Generating Bounding Box Supervision for Semantic Segmentatio...

arXiv 2016
[7023]

Adversarial synthesis learning enables segmentation without target modality ground truth, in: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)

https://doi.org/10.1109/CVPR.2018.00733 Huo, Y ., Xu, Z., Bao, S., Assad, A., Abramson, R.G., Landman, B.A., 2018. Adversarial synthesis learning enables segmentation without target modality ground truth, in: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). Presented at the 2018 IEEE 15th International Symposium on Biomedical Imag...

work page arXiv 2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [53]

https://doi.org/10.1093/nsr/nwx106

doi:10.1093/nsr/nwx106

[2] [191]

Why Should I Trust You?

https://doi.org/10.1007/978-3-030-00889-5_21 Pan, Y ., Liu, M., Lian, C., Zhou, T., Xia, Y ., Shen, D., 2018. Synthesizing Missing PET from MRI with Cycle-consistent Generative Adversarial Networks for Alzheimer ’s Disease Diagnosis, in: Frangi, A.F., Schnabel, J.A., Davatzikos, C., Alberola-López, C., Fichtinger, G. (Eds.), Medical Image Computing and Co...

arXiv 2018

[3] [2600]

Unsupervised Domain Adaptation in Brain Lesion Segmentation with Adversarial Networks, in: Niethammer, M., Styner, M., Aylward, S., Zhu, H., Oguz, I., Yap, P

https://doi.org/10.1109/TCSVT.2016.2589879 Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V ., Simpson, J., Kane, A., Menon, D., Nori, A., Criminisi, A., Rueckert, D., Glocker, B., 2017a. Unsupervised Domain Adaptation in Brain Lesion Segmentation with Adversarial Networks, in: Niethammer, M., Styner, M., Aylward, S., Zhu, H., Oguz, I., Yap, P. -T.,...

arXiv 2016

[4] [2854]

Nutrition and Care of the Preterm Infant: Current Issues

https://doi.org/10.1109/CVPR.2016.311 Bloch, N., Madabhushi, A., Huisman, H., 2015. NCI-ISBI 2013 Challenge: Automated Segmentation of Prostate Structures. The Cancer Imaging Archive 370. https://doi.org/10.7937/k9/tcia.2015.zf0vlopv Bonechi, S., Andreini, P., Bianchini, M., Scarselli, F., 2018. Generating Bounding Box Supervision for Semantic Segmentatio...

arXiv 2016

[5] [7023]

Adversarial synthesis learning enables segmentation without target modality ground truth, in: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)

https://doi.org/10.1109/CVPR.2018.00733 Huo, Y ., Xu, Z., Bao, S., Assad, A., Abramson, R.G., Landman, B.A., 2018. Adversarial synthesis learning enables segmentation without target modality ground truth, in: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). Presented at the 2018 IEEE 15th International Symposium on Biomedical Imag...

work page arXiv 2018