PaperlayerView exact revision on arXiv

Source-linked AI summary

Expert-Guided Explainable Few-Shot Learning with Active Sample Selection for Medical Image Analysis

Longwei Wang, Ifrat Ikhtear Uddin, KC Santosh

arXiv:2601.02409v1eess.IVcs.AIcs.CV

TL;DR

Medical image analysis faces scarce labeled data and limited interpretability in few-shot models and active learning. The paper combines expert-guided few-shot learning with explainability-guided sample acquisition. Across MRI and chest X-ray tasks, the conclusion reports classification gains and improved data efficiency over random sampling.

  • Problem

    Medical imaging has scarce expert-labeled data, while few-shot models often lack interpretability and active learning overlooks the interpretability of selected samples.

    Source: §I · p. 1

  • Method

    EGxFSL uses radiologist annotations with Grad-CAM-based explanation supervision alongside prototypical classification, while xGAL selects samples using uncertainty and explanation misalignment.

    Source: §I · pp. 1–2

  • Results

    Across MRI and chest X-ray modalities, EGxFSL improved classification by 3–20 percentage points; xGAL achieved competitive performance with 680 labeled samples and 15–38% relative improvements over random sampling.

    Source: §VI · p. 13

  • Takeaways & Limitations

    The authors describe the combination of interpretability and data efficiency as particularly suitable for resource-constrained clinical environments, rare disease diagnosis, and specialized imaging protocols.

    Source: §VI · p. 13

Abstract

from arXiv · show

Medical image analysis faces two critical challenges: scarcity of labeled data and lack of model interpretability, both hindering clinical AI deployment. Few-shot learning (FSL) addresses data limitations but lacks transparency in predictions. Active learning (AL) methods optimize data acquisition but overlook interpretability of acquired samples. We propose a dual-framework solution: Expert-Guided Explainable Few-Shot Learning (EGxFSL) and Explainability-Guided AL (xGAL). EGxFSL integrates radiologist-defined regions-of-interest as spatial supervision via Grad-CAM-based Dice loss, jointly optimized with prototypical classification for interpretable few-shot learning. xGAL introduces iterative sample acquisition prioritizing both predictive uncertainty and attention misalignment, creating a closed-loop framework where explainability guides training and sample selection synergistically. On the BraTS (MRI), VinDr-CXR (chest X-ray), and SIIM-COVID-19 (chest X-ray) datasets, we achieve accuracies of 92\%, 76\%, and 62\%, respectively, consistently outperforming non-guided baselines across all datasets. Under severe data constraints, xGAL achieves 76\% accuracy with only 680 samples versus 57\% for random sampling. Grad-CAM visualizations demonstrate guided models focus on diagnostically relevant regions, with generalization validated on breast ultrasound confirming cross-modality applicability.

I. INTRODUCTION

Medical imaging FSL and active learning address scarce labels, but conventional approaches leave model transparency or interpretability-aware sample selection unresolved. The paper proposes EGxFSL and xGAL to combine expert-guided attention with data-efficient learning.

A. FSL in Medical Imaging

  • Few-shot learning can generalize from a handful of labeled instances per class, but most FSL models lack interpretability, leaving their decisions opaque to clinicians.Prototypical networks form class prototypes in an embedding space.Source: §I · p. 1, A. FSL in Medical Imaging · p. 2

B. AL for Annotation Efficiency

  • Traditional active learning selects informative unlabeled samples, but often relies on unreliable softmax uncertainty and does not consider whether selected samples improve interpretability.Medical AL work also includes entropy- and diversity-based selection and query-by-committee strategies.Source: §I · p. 1, B. AL for Annotation Efficiency · p. 2

I. INTRODUCTION

  • Experiments on BraTS, VinDr-CXR, and SIIM-FISABIO-RSNA-COVID-19 report that EGxFSL and xGAL outperform baseline FSL and traditional AL in accuracy and interpretability metrics while reducing labeled-data needs.The contribution is evaluated on three expert-annotated datasets.Source: §I · p. 2
  • EGxFSL aligns Grad-CAM heatmaps with radiologist ROIs using a Dice-based explanation loss, while xGAL selects samples using predictive uncertainty and explanation misalignment.xGAL uses EGxFSL as its core learner, and the framework is validated across three datasets and two imaging modalities.Source: §I · pp. 1–2

C. xAI in Medical Imaging

Post-hoc explanations may look plausible without influencing learning, while explanation-guided methods have not addressed medical few-shot learning and active sample selection together. This section motivates and outlines an expert-guided prototypical approach.

C. xAI in Medical Imaging

  • Post-hoc Grad-CAM and SHAP visualize salient regions, but because they are applied after training, models may still rely on spurious features despite plausible explanations.The passage connects this disconnect to reduced clinical reliability.Source: C. xAI in Medical Imaging · p. 2

1) Prototypical Network Foundation:

  • The framework diagram shows support-set prototypes classifying queries, followed by query Grad-CAM generation and Dice alignment with expert annotations during joint optimization.The described backbone is DenseNet-121.Source: Figure 1 · p. 3

D. Explanation-Guided Learning and Supervised Attention Alignment

  • Prior explanation-guided work includes LRP for glioma MRI and radiologist ROI guidance for lung-cancer CT, but does not combine medical FSL, sample efficiency, and clinical interpretability.The cited work also leaves a gap in unifying explanation-guided training with active learning under data-scarce conditions.Source: D. Explanation-Guided Learning and… · p. 3

III. METHODOLOGY

  • The proposed dual framework pairs ROI-supervised FSL using Grad-CAM Dice loss with active learning that selects samples using predictive uncertainty and attention misalignment.The framework integrates radiologist-provided spatial annotations into both model training and sample selection.Source: D. Explanation-Guided Learning and… · p. 3, §III · p. 3

A. EGxFSL: Expert-Guided Explainable Few-Shot Learning

2) Expert-Guided Attention Alignment via Grad-CAM:

EGxFSL compares query Grad-CAM maps with expert abnormality masks and combines explanation alignment with classification training. xGAL extends the approach by selecting unlabeled samples using both prediction and attention criteria.

2) Expert-Guided Attention Alignment via Grad-CAM:

  • Grad-CAM maps for query images are compared with expert binary masks of abnormality locations using a Dice-based explanation alignment loss.The loss increases when attention is misaligned and supplies gradients to correct model focus during backpropagation.Source: 2) Expert-Guided Attention Alignment… · p. 4

3) Joint Optimization Objective:

  • The total EGxFSL objective combines prototypical classification loss with explanation loss, with α controlling the strength of explanation-based supervision.Minimizing the combined loss encourages both classification accuracy and clinically aligned attention.Source: 3) Joint Optimization Objective · p. 4

B. xGAL: Explainability-Guided Active Learning

1) Scoring Mechanism for Sample Acquisition:

xGAL scores unlabeled samples using classification uncertainty and explanation misalignment, then labels selected samples and retrains EGxFSL in repeated rounds. The acquisition process uses both class labels and expert spatial annotations.

1) Scoring Mechanism for Sample Acquisition:

  • Shannon entropy over the model’s softmax output measures classification uncertainty, with higher entropy indicating a potentially valuable sample to label.Explanation misalignment compares the predicted-class Grad-CAM map with the expert-supplied diagnostic region using a Dice-based loss.Source: 1) Scoring Mechanism for Sample… · p. 4
  • xGAL combines classification uncertainty and Grad-CAM misalignment with expert regions in a composite score for ranking unlabeled samples.The balance parameter λ controls the two criteria; reported settings are 0.5 for BraTS and SIIM-COVID and 0.6 for VinDr-CXR.Source: 1) Scoring Mechanism for Sample… · p. 4
  • At each round, the top-K samples are selected and annotated with class labels and expert masks, represented as segmentation maps for BraTS or bounding-box-derived binary masks for VinDr-CXR and SIIM-COVID.Source: 2) Active Learning Procedure · p. 5

2) Active Learning Procedure:

  • Selected samples join the labeled support set, and EGxFSL retrains on the expanded set; the cycle repeats, with experiments reporting improvements in accuracy and attention alignment using 680 labeled samples.The procedure is described as iteratively refining classification and interpretability.Source: 2) Active Learning Procedure · p. 5, 3) Algorithm Summary · p. 5

3) Algorithm Summary:

The algorithm combines uncertainty and explanation misalignment to select samples for EGxFSL retraining, while attribution guidance is presented as a constraint on few-shot hypotheses. The section also gives dataset-specific guidance for choosing supervision and acquisition weights.

C. Hyperparameter Selection Guidelines

  • The xGAL algorithm computes entropy and Grad-CAM misalignment for unlabeled samples, scores them as λH(x)+(1−λ)Dexp(x), selects the top K, and fine-tunes on the expanded labeled set.The procedure then removes selected samples from the unlabeled pool.Source: C. Hyperparameter Selection Guidelines · p. 5
  • The recommended explanation-supervision strength is α = 0.10 across datasets, with 0.20–0.30 suggested for poor attention alignment and 0.05 if classification accuracy degrades.The passage recommends validating choices over a grid from 0.05 to 0.50.Source: C. Hyperparameter Selection Guidelines · p. 5
  • The acquisition balance is λ = 0.5 for balanced BraTS and SIIM-COVID datasets and λ = 0.6 for imbalanced VinDr-CXR, with higher values favoring accuracy and lower values favoring interpretability.The guidance recommends starting at 0.5 and refining against validation performance.Source: C. Hyperparameter Selection Guidelines · pp. 5–6

A. Attribution-Guided Supervision Reduces Hypothesis Space Overfitting

  • Few-shot learning has a large hypothesis space relative to its labeled examples, while prototypical prediction loss does not constrain internal feature importance or spatial saliency.The passage notes that models may consequently exploit spurious correlations or background patterns and memorize rather than generalize.Source: A. Attribution-Guided Supervision… · p. 6
  • Dice-based explanation loss aligns model Grad-CAM heatmaps with expert masks, constraining attention to clinically meaningful regions and supporting compact, clinically relevant representations.Source: A. Attribution-Guided Supervision… · p. 6
  • The attribution-consistent hypothesis class is a subset of the predictive class, imposing a threshold τ on interpretability alignment and reducing effective capacity.Source: A. Attribution-Guided Supervision… · p. 6
  • An information-bottleneck view describes representations as predictive of labels and informative about expert attention maps while constrained to discard irrelevant input details.Source: A. Attribution-Guided Supervision… · p. 6

B. Formal Bounds on Expected Gain in Active Sample Selection using Mutual Information

This section frames active sample selection through mutual information about labels, model parameters, explanations, and expert annotations. It relates those objectives to the uncertainty and misalignment terms used by xGAL.

  • In data-constrained medical imaging, mutual information is presented as a measure of the expected utility of labeling an unlabeled sample through its effect on uncertainty about model parameters or predictions.Source: B. Formal Bounds on Expected Gain in… · p. 6
  • BALD selects samples by mutual information between predicted labels and model parameters, expressed as predictive entropy minus expected single-model entropy.The passage characterizes high values as epistemic uncertainty when the ensemble is uncertain but individual models make confident, differing predictions.Source: B. Formal Bounds on Expected Gain in… · p. 6
  • The framework defines explanation-related mutual information between an explanation map and the expert diagnostic region; both are spatial masks that can be estimated with pixel-wise Bernoulli models.The passage also mentions discretization or variational approximations for estimation.Source: B. Formal Bounds on Expected Gain in… · p. 6
  • xGAL uses entropy as a tractable proxy for label-parameter mutual information and Dice-based misalignment as a proxy for the complement of explanation-annotation mutual information.Source: B. Formal Bounds on Expected Gain in… · p. 6

V. EXPERIMENTS AND RESULTS

The experiments cover expert-annotated MRI and chest X-ray datasets, with classification and interpretability measures specified. The described EGxFSL evaluation uses episodic few-shot configurations and a DenseNet-121 feature extractor.

A. Datasets

  • The evaluation datasets are BraTS multimodal brain-tumor MRI, VinDr-CXR frontal chest X-rays, and SIIM-COVID chest radiographs.BraTS includes expert tumor-subregion segmentation masks; VinDr-CXR uses bounding boxes for three thoracic findings, and SIIM-COVID has three COVID-19 appearance classes.Source: A. Datasets · p. 7

B. Evaluation Metrics

  • Classification is measured by overall accuracy and macro AUC, while interpretability is assessed by comparing Grad-CAM maps with expert ROIs and by qualitative visualizations.Source: B. Evaluation Metrics · p. 7

C. Experiments on EGxFSL

  • EGxFSL uses episodic 3-way K-shot training, evaluates 1-shot, 3-shot, and 5-shot settings, and uses DenseNet-121 as its feature extractor.The passage introduces a comparison but does not specify its candidates or results.Source: 1) Training Protocol · p. 7

1) Training Protocol:

The protocol compares prototypical networks with and without expert-guided explanation supervision. EGxFSL outperforms the non-guided baseline across the reported few-shot configurations and datasets, with α = 0.10 yielding the best accuracy in the sensitivity analysis.

1) Training Protocol:

  • The guided model adds explanation supervision to the prototypical loss, while the non-guided baseline uses only the prototypical loss.The guided model was trained with Ltotal = Lproto + α · Lexp, and α = 0.10 was selected through validation.Source: 1) Training Protocol · p. 7

2) Performance Evaluation:

  • The guided model’s best 5-shot VinDr-CXR result was 76.12% accuracy and 87.29% macro AUC (10 epochs, 100 episodes), while SIIM-COVID reached 62.08% and 80.00%, respectively (10 epochs, 150 episodes).Source: 2) Performance Evaluation · p. 7
  • Grad-CAM comparisons show the guided model focusing on expert-annotated clinically relevant regions, unlike the non-guided model’s irrelevant attention.The comparison covers BraTS, VinDr-CXR, and SIIM-COVID.Source: Figure 4 · p. 8

3) Effect of Explanation Weight

  • At α = 0.10, accuracy peaked across BraTS, VinDr-CXR, and SIIM-COVID; both lower and higher values degraded performance.In the 3-shot, 7-epoch, 60-episode setup, the reported accuracies were 83.61%, 73.29%, and 55.83%, respectively.Source: Figure 3 · p. 7, 3) Effect of Explanation Weight · p. 9

2) Performance Evaluation:

  • The guided model improved class-wise F1 across diagnostic categories, including VinDr-CXR Nodule/Mass from 51.04% to 75.09% and Pulmonary Fibrosis from 42.05% to 74.08%.The passage reports consistent gains across all classes on BraTS, VinDr-CXR, and SIIM-COVID.Source: Table II · p. 8, 4) Class-wise Performance Analysis · p. 9

4) Class-wise Performance Analysis:

xGAL iteratively selects samples using both predictive uncertainty and explanation misalignment, then retrains the EGxFSL learner. Its reported 5-shot results exceed random-sampling results across the three datasets, and the best acquisition balance varies by dataset.

D. Experiments on xGAL

  • xGAL starts from a 200-sample EGxFSL baseline, selects 120 samples in each of four rounds, and retrains after each round, totaling 680 samples.The random-sampling comparison uses the same total number of labeled samples.Source: 1) xGAL Setup · p. 9, D. Experiments on xGAL · p. 9

2) Training Sample Acquisition:

  • At each iteration, xGAL scores unlabeled samples using classification entropy and Grad-CAM-to-expert-annotation Dice misalignment, then uses the retrained model for the next round.The acquisition score balances uncertainty and explanation misalignment with λ.Source: 2) Training Sample Acquisition · p. 9

3) Impact of Acquisition Weight

4) AL Results:

  • The 5-shot xGAL results included macro AUCs of 90.00% on BraTS, 68.21% on VinDr-CXR, and 66.92% on SIIM-COVID.For VinDr-CXR and SIIM-COVID, the reported accuracy comparisons were 52.37% versus 45.49% and 52.66% versus 38.28% for xGAL versus random sampling.Source: 4) AL Results · pp. 9–10
  • 76.85% accuracy on BraTS, 52.37% on VinDr-CXR, and 52.66% on SIIM-COVID were the reported xGAL 5-shot results, exceeding random sampling on each dataset.The corresponding random-sampling accuracies were 58.01%, 45.49%, and 38.28%, respectively, using 680 total samples.Source: 4) AL Results · pp. 9–10, Table III · p. 10

E. Interpretability Analysis

The interpretability analysis reports closer alignment between EGxFSL attention and expert-marked regions, with more concentrated focus in higher-shot settings. The section also reports multi-seed validation and introduces an ablation study of explanation alignment and acquisition choices.

E. Interpretability Analysis

  • Guided Grad-CAM maps focus on diagnostically meaningful regions, while non-guided maps often emphasize irrelevant areas; 3-shot and 5-shot focus is more concentrated than 1-shot focus.The authors state that additional support samples improve alignment with expert-defined areas.Source: E. Interpretability Analysis · p. 10

F. Statistical Significance Validation

  • For EGxFSL, the guided model achieved 91.6% ± 2.1% accuracy versus 89.2% ± 2.8% baseline, with macro AUC 96.6% ± 1.3% versus 95.5% ± 1.7% (p = 0.040).The overall accuracy p-value was at the significance threshold (p < 0.050); Cohen’s d values were 0.73–0.98.Source: F. Statistical Significance Validation · p. 10
  • With 680 samples, xGAL achieved 76.85% ± 4.1% accuracy versus 56.5% ± 1.3% for random sampling (p = 0.001).Macro AUC was 88.6% ± 2.7% versus 71.4% ± 0.3% (p < 0.001).Source: F. Statistical Significance Validation · p. 10

G. Ablation Study

1) Effect of Explanation Alignment Loss:

The ablations compare expert-grounded attention supervision with baseline and noisy supervision, then compare active-learning criteria and attribution methods. The reported findings favor clinically grounded alignment, a dual acquisition objective, and Grad-CAM’s balance of accuracy, interpretability, and training time.

1) Effect of Explanation Alignment Loss:

  • The guided model outperformed baseline and randomly permuted attention supervision in both classification accuracy and attention-map IoU.The authors report that the random-supervision variant suffered on both metrics, supporting the value of grounding alignment in clinical knowledge.Source: 1) Effect of Explanation Alignment Loss · p. 11

2) Effect of AL Acquisition Strategy:

  • With the same 680-sample budget, dual-objective xGAL outperformed the acquisition baselines on accuracy and attention alignment; Dice-only performed better than entropy-only.The dual objective combines entropy and explanation misalignment with λ = 0.5.Source: 2) Effect of AL Acquisition Strategy · p. 11, Table VI · p. 11

3) Effect of Attribution Method for Explanation Supervision:

  • Grad-CAM achieved the best balance of accuracy, interpretability, and training time; Integrated Gradients was comparable but costlier, while SHAP was noisy and slow.These attribution methods were compared on BraTS.Source: 3) Effect of Attribution Method for… · p. 11, Table VII · p. 11
  • On breast ultrasound, EGxFSL achieved 87.6% accuracy and 94.5% macro AUC, with attention aligned to expert-marked lesion boundaries.The BUSI dataset contained 600 normal, benign, and malignant images; the reported result supports applicability beyond radiology.Source: 4) Generalization to Ultrasound Imaging · p. 11

4) Generalization to Ultrasound Imaging:

The discussion treats expert ROIs as probabilistic guidance, while the reported visualizations show guided attention aligning with expert regions across ultrasound, MRI, and chest X-ray examples.

4) Generalization to Ultrasound Imaging:

  • On BUSI ultrasound images, guided attention aligns with lesion boundaries, whereas non-guided attention is misaligned, supporting generalization beyond MRI and X-ray.The examples include benign and malignant lesions.Source: Figure 7 · p. 12

H. Discussion

  • Grad-CAM explanation alignment acts as a soft spatial regularizer, encouraging consistency with expert ROIs without treating annotations as absolute ground truth.This probabilistic treatment is intended to accommodate incomplete, noisy, or inconsistent annotations.Source: 1) Annotation Variability and Clinical… · p. 12

1) Annotation Variability and Clinical Reliability:

  • xGAL prioritizes informative, diagnostically uncertain cases for review; comparable diagnostic performance with 680 labeled samples reduces manual annotation workload.The framework also treats radiologist annotation variability as part of its probabilistic learning process.Source: 2) Clinical Applicability and… · pp. 12–13

2) Clinical Applicability and Integration into Workflow:

  • The workflow combines selective review with soft supervision that steers attention toward expert-indicated regions without overfitting to noisy or incomplete labels.The paper describes the components as designed for annotation cost, inter-radiologist variability, and annotation-quality constraints.Source: 2) Clinical Applicability and… · p. 13
  • Evaluation is limited to MRI, chest X-ray, and preliminary ultrasound; the approach also requires spatial annotations and faces Grad-CAM resolution and few-shot noise constraints.The paper reports stronger performance with fine-grained segmentation masks than with coarse bounding boxes.Source: 3) Limitations and Future Work · p. 13

VI. CONCLUSION

  • EGxFSL integrated radiologist-provided regions of interest through a Grad-CAM-based Dice loss, improving classification by 3–20 percentage points and aligning attention clinically across MRI and chest X-ray modalities.Source: §VI · p. 13

Related papers

Loading 2601.02409v1…