A Systematic Review of Few-Shot Learning in Medical Imaging

1.1. Rationale

• Medical datasets are often small because of privacy concerns, high acquisition costs, and laborious expert annotation.Few-shot learning is presented as relevant to rare conditions, new tasks, and patient-specific analysis where examples may be limited.Source: §1.1 · p. 1

1.2. Objectives

• The review addresses limited existing medical-imaging FSL reviews, which cover few papers and outcomes and omit statistical synthesis, bias analysis, and shared-method discussion.Source: §1.2 · p. 2
• The authors aim to identify substantial studies across segmentation, classification, and registration, assess risk of bias and applicability, and establish a reusable methodological pipeline.Source: §1.2 · p. 2
• The review analyzes study distributions and results across outcomes, clinical tasks, and meta-learning techniques.It also examines data use, imaging modalities, and model-robustness assessment methods.Source: §1.2 · p. 2

• Meta-learning trains across tasks or episodes to acquire transferable knowledge for predictions on new tasks with few examples.The framework uses an inner algorithm for task-level learning and an outer algorithm to update parameters across tasks.Source: §2 · p. 3
• Meta-training samples source tasks with support and query sets, while meta-testing adapts on unseen target tasks; FSL typically uses fewer than 10 examples per class in the support set.Source: §2 · p. 3
• ZSL concerns test samples from unseen classes, while GZSL allows both seen and unseen classes; attribute transfer is a common approach.Other approaches include intermediate attribute classifiers, seen-class proportions, and compatibility functions.Source: §2 · p. 3
• The review categorizes FSL meta-learning as initialization-based, metric learning-based, or hallucination-based, reflecting different forms of meta-knowledge.Source: §2 · p. 3

2.1. Initialization-based methods

• Initialization-based methods learn parameters that can be fine-tuned with few examples from a new episode for rapid generalization.Source: §2.1 · p. 3

2.1.1. Model-Agnostic Meta-Learning

• MAML learns parameters that adapt to new tasks through gradient descent, while FOMAML omits second derivatives and performed almost as well in the reported results.The proposed first-order approximation addresses the computational expense of gradients for task and meta objectives.Source: §2.1.1 · p. 4

2.1.2. Reptile

• Reptile [19] updates global parameters using task-specific parameters, treating their difference as a gradient for the final update.The review describes this update as more computationally efficient than MAML's second-order differentiation.Source: §2.1.2 · p. 5

2.1.3. Optimization as Long Short-Term Memory network cell update

• The LSTM optimizer receives task gradients and losses and generates updated model parameters, learning an optimization rule from the analogy between gradient updates and LSTM cell updates.The updated parameters are evaluated on the test set after a predefined number of steps, and that loss trains the meta-learner.Source: §2.1.3 · pp. 5–6

2.1.4. Optimization with Markov decision process and Reinforcement Learning

• The reinforcement-learning approach [22] represents optimization as an MDP and learns a policy that maps objective values and gradients along an optimizer trajectory to the next step vector.Its convergence speed is used as the cost function, penalizing policies that converge slowly.Source: §2.1.4 · p. 6

2.1.5. Memory-augmented Neural Networks

• Neural Turing Machines [25], a differentiable form of MANNs, use a controller and external memory to retrieve labels associated with previously seen inputs.Inputs and labels are presented one time step apart and shuffled within each episode to prevent direct label-output mapping or sequence learning.Source: §2.1.5 · p. 6

2.2.1. Siamese Neural Networks

• Siamese networks use two weight-sharing networks to compare image pairs, then classify a novel image by its highest-scoring match.The networks learn to distinguish same-class from different-class pairs and are trained on many classes.Source: §2.2.1 · pp. 6–7

2.2.2. Triplet Networks

• Triplet Networks learn from anchor, positive, and negative samples, outputting L2 distances that support same-class decisions at inference.A related Dual-Triplet Network learns visual-semantic mappings for zero-shot classification using nearest-neighbour search on attributes.Source: §2.2.2 · p. 7

2.2.3. Matching Networks

• Matching Networks map a support set to a classifier that produces a query-label distribution using attention over support examples.The simplest attention implementation applies softmax to cosine distance, and training uses support/query episodes.Source: §2.2.3 · p. 7

2.2.4. Prototypical Networks

• Prototypical Networks average support-sample embeddings into class prototypes and classify queries by their distances to those prototypes.They use negative squared Euclidean distance here; the authors state that they are equivalent to Matching Networks in one-shot learning, and metadata embeddings can serve as prototypes for zero-shot learning.Source: §2.2.4 · pp. 7–8

2.2.5. Relation Networks

• Relation Networks learn the comparison metric end-to-end, scoring concatenated support and query embeddings with a relation module.For generic few-shot learning, class feature maps sum embedding outputs; for zero-shot learning, separate modules embed semantic attributes and images.Source: §2.2.5 · p. 8

2.3. Hallucination-based methods

• Hallucination-based methods generate additional data to address scarcity in the available dataset.The section introduces two approaches: learning transformations from intra-class analogies and jointly training a hallucinator with a classifier.Source: §2.3 · p. 8, §2.3.1 · p. 8, §2.3.2 · p. 9

2.3.1. Hallucinating with Intra-class Analogies

• Intra-class analogies train a transformation on base-class examples and apply it to novel-class images to generate hallucinated features.The transformation takes three feature vectors as input and is implemented as a three-layer MLP.Source: §2.3.1 · p. 8

2.3.2. Classificator and Hallucinator End-to-End Model

• The end-to-end model jointly trains a hallucinator and classifier, generating extra examples to augment each sampled meta-training set [36].At meta-test time, hallucinated examples are added to the initial training set before classifier training; the approach is agnostic to the specific meta-learning algorithm.Source: §2.3.2 · p. 9

3.1. Study Design

• The review aims to establish a shared methodological pipeline and analyze study distributions, performance, and imaging modality across segmentation, classification, and registration.The study design follows the PRISMA 2020 checklist.Source: §3.1 · p. 9

3.2. Eligilibity criteria

• Eligible papers had to implement few-shot learning in medical imaging and use at most 20 training examples per class.The review excluded abstracts, non-peer-reviewed or non-English papers, papers with significant theoretical errors, and few-shot domain adaptation studies.Source: §3.2 · p. 9

3.3. Information sources

• The review used databases as information sources for its literature search.The passage introduces the databases but does not name them.Source: §3.3 · p. 9
• Table 1 lists the research queries employed for each database.Source: Table 1 · p. 10
• Searches were conducted on September 7, 2022, and January 25, 2023; NILDE was used to request inaccessible full texts.Source: §3.3 · p. 10

3.4. Research strategies

• The review listed the search queries used for each database in Table 1.Source: §3.4 · p. 10

3.5. Selection process

• A single reviewer manually screened records, with no machine-learning assistance, crowdsourcing, or prescreened datasets.Source: §3.5 · p. 10

3.6. Data collection process

• One reviewer manually extracted relevant information from each report, without automated collection or figure-extraction tools.Source: §3.6 · p. 10

3.7. Data item

• The review examined segmentation, classification, and registration, alongside use of FSL, OSL, and ZSL techniques in medical imaging.Source: §3.7 · p. 11

3.8. Assessment of bias in studies

• PROBAST was used to assess risk of bias and applicability, with tables marking low concerns by checkmarks and high concerns by crosses.Source: §3.8 · p. 11

3.9. Effect measures

• Segmentation was assessed with Dice score and Intersection over Union, classification with Accuracy, F1-score, Recall, and AUROC, and registration with Dice score, ALD, and TRE.Source: §3.9 · p. 11

3.10. Synthesis methods

• Forest plots grouped each study’s highest metric performance by clinical task or meta-learning method and displayed group means and 95% confidence intervals, but no meta-analysis was conducted.The review cited variation across clinical applications as making direct comparisons inappropriate; the plots therefore visualized individual study findings rather than quantitatively synthesizing them.Source: §3.10 · p. 11
• The review also tabulated learning strategies across pre-training, training, and data augmentation to identify a shared methodological pipeline.Source: §3.10 · p. 11

4.1. Study selection

• 314 studies were retrieved, and 80 were included in the final analysis, as summarized by the PRISMA flow diagram.Source: §4.1 · p. 11, Figure 4 · p. 12

4.2. Studies characteristics

• Studies were analyzed across three outcomes, and,,,, and were counted multiple times because they address multiple outcomes.Source: §4.2 · p. 11, Figure 5 · p. 13

4.2.1. Segmentation

• Among 50 studies focused primarily on medical segmentation, approaches included implicit MAML for same- and different-task evaluation, semi-supervised coarse-to-fine GAN segmentation, and pretraining or registration-based augmentation.Xu and Niethammer used a pre-trained network and joint registration-segmentation training, with registration providing realistic data augmentation.Source: §4.2.1 · pp. 11–13
• Segmentation most often targeted liver and kidney (18 studies each), spleen (17), heart (14), and brain (12), among many less frequent anatomical tasks.These counts correspond to 36% each for liver and kidney, 34% for spleen, 28% for heart, and 24% for brain; Figure 6a depicts the leading tasks.Source: §4.2.1 · p. 22, Figure 6 · p. 22
• Of 50 segmentation studies, 28 (56%) used no meta-learning; metric-based methods appeared in 14 (28%), initialization in six (12%), and hallucination methods in three (6%).For shot configuration, 20 studies (40%) focused exclusively on OSL, 15 (30%) used k-shot training with k from 2 to 20, 14 (28%) combined OSL and FSL, and one (2%) used ZSL.Source: §4.2.1 · p. 22, Figure 6 · p. 22
• Segmentation studies used MRI (30, 60%) and CT (26, 52%) most often; 21 (42%) conducted ablation studies only, while 13 (26%) used no specific evaluation technique.Eleven studies (22%) combined ablation and cross-validation, and five (10%) used cross-validation alone.Source: §4.2.1 · p. 22
• In segmentation pipelines, 33 of 50 studies (66%) omitted pre-training; main training most often used meta-learning (20, 40%), while 24 (48%) used no data augmentation.Supervised learning was used for pre-training in 13 studies (26%), and classical augmentation was the most frequent augmentation type (16, 32%).Source: §4.2.1 · p. 23, Table 4 · p. 26

4.2.2. Classification

• Among 27 classification studies, PFEMed uses episodic dual encoders and a prior-guided VAE, meta-trained models are fine-tuned with supervised learning, and SSL supports generalized zero-shot learning.The review identifies the 27 studies as primarily focused on medical classification.Source: §4.2.2 · p. 27
• Classification results include Accuracy: 0.851 (Brain, MAT), 0.960 (Skin, MAT), 0.803 (Cervix, MAT), and 0.762 (Cells, MAS), 0.762 (Colon, MAD), 0.506 (Lungs, MAD).One-shot AUROC was 0.961 (Eye, MAD) and 0.955 (Lung, COVID); another reported Accuracy was 0.346 (Lungs, MAD).Source: §4.2.2 · p. 29
• Reported classification accuracies were 0.921 for Breast LN metastases, 0.909 for Lungs MAD, 0.942 for Eye DE, and 0.911 for Prostate tumour; 4-shot AUROC was 0,910 for Breast tumour.Source: §4.2.2 · p. 29
• Other reported classification results include AUROC: 0.679 (Lung, MAD), Accuracy: 0.857 (Lung, COVID), and Accuracy/AUROC pairs of 0.819/0.843 (Skin, MAD) and 0.703/0.734 (Lungs, MAD).The same reported outcomes include F1-score: 0.440 (Lung, MAD) and Recall: 0.490 (Lung, MAD).Source: §4.2.2 · p. 29
• Lung classification was the most common clinical task (11/27 studies, 41%), and metric-learning methods were most frequent (10/27, 37%).Breast and skin classification appeared in 22% and 26% of studies, respectively; 11 studies (41%) used no meta-learning, while six (22%) used initialization-based methods. Training setups, modalities, and evaluation approaches varied, and Figures 10–11 summarized mean and 95% CI performance by task and meta-learning method.Source: §4.2.2 · pp. 32–33
• Classifiers most often used classical supervised pre-training (13/27, 48%) and meta-learning for training (15/27, 56%).Classical augmentation was used by 10 studies (37%); 15 (56%) used no augmentation. Table 7 outlines the pipeline steps adopted by classification studies.Source: §4.2.2 · p. 34, Table 7 · p. 34

4.2.3. Registration

• All six registration studies omitted meta-learning, and four (67%) used unsupervised training without pre-training.Brain, heart, and lung registration each featured in three studies (50%). MRI was used in five studies (83%) and CT in four (67%); two studies (33%) used supervised approaches for both pre-training and training. Table 10 outlines registration pipeline steps.Source: §4.2.3 · pp. 36–37, Table 10 · p. 38

4.3. Standard pipeline

• The review proposed a standardized few-shot learning pipeline based on its analysis of classification, segmentation, and registration studies.Figure 15 illustrates the pipeline and highlights learning frameworks suited to pre-training and training.Source: §4.3 · p. 38, Figure 15 · p. 38

5. Discussion

• Segmentation accounted for 61% of reviewed studies, classification 32%, and registration 7%; heart, abdomen, and lungs were primary application areas.The heart was most investigated in segmentation and registration, while lungs led classification.Source: §5 · pp. 39, 42
• Mean Dice scores were 0.89 for AA and 0.88 for LV segmentation, while lung segmentation had an IoU of 0.91 across several studies.The review notes that its findings draw on a limited pool of studies, potentially affecting generalizability; femur’s highest Dice score came from one study.Source: §5 · p. 39
• No-meta-learning and hallucination-based methods had the highest mean segmentation Dice score (0.84), while metric-learning methods averaged 0.79.For classification, no-meta-learning studies achieved the highest accuracy and AUROC; registration studies used non-meta-learning approaches.Source: §5 · pp. 39–40
• Supervised learning was the most common pre-training approach, while meta-learning dominated training for segmentation and classification; registration studies used unsupervised or weakly supervised training.Only a few studies used pre-training, and the review recommends exploring self-supervised learning and meta-learning during pre-training.Source: §5 · pp. 40–41
• Supervised learning was the second most employed approach for segmentation and classification, while self-supervised, zero-shot, and semi-supervised methods also appeared.The review recommends further exploration of self-supervised and semi-supervised approaches, including combinations with meta-learning.Source: §5 · p. 41
• The reviewed augmentation methods included classical, registration-based, and generative approaches, including atlas-based synthesis of images using shape and intensity changes.Registration-based augmentation included using registration outputs as additional training data; generative examples included pasted lesion patches and synthesized images.Source: §5 · pp. 41–42
• Non-meta-learning methods outperform meta-learning approaches on average, but class mismatch and cross-task generalizability complicate this comparison.The review argues that models trained on diverse tasks may encounter unseen classes at test time, so average performance alone does not capture generalizability.Source: §5 · p. 42
• The review excluded image-reconstruction studies and may be incomplete because it omitted medical databases.It recommends broader medical applications, meta-learning for registration, and thorough model validation; some classification and registration studies lacked comprehensive model analyses.Source: §5 · p. 42

6. Conclusions

• The review analyzes state-of-the-art few-shot learning results by clinical task and meta-learning algorithm.These analyses culminate in a standard methodological pipeline shared across the examined studies.Source: §6 · pp. 42–43
• The authors present a standard methodological pipeline shared across the examined studies.They intend the pipeline to guide further research.Source: §6 · p. 43
• They advocate self-supervised, semi-supervised, and meta-learning approaches, broader clinical-domain research, and robustness and generalizability assessment.The recommendations also include exploring hallucination-based methods for few-shot learning in medical imaging.Source: §6 · p. 43