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

TL;DR

The paper tackles the twin challenges of limited labeled data and lack of interpretability in medical image analysis by proposing a dual-framework: Expert-Guided Explainable Few-Shot Learning (EGxFSL) and Explainability-Guided Active Learning (xGAL). EGxFSL integrates radiologist-defined ROIs via Grad-CAM-based Dice supervision into a prototypical network, yielding accurate few-shot classification with clinically aligned attention, as demonstrated across BraTS, VinDr-CXR, and SIIM-COVID. xGAL further enhances data efficiency by selecting unlabeled samples based on a combined criterion of predictive uncertainty and explanation misalignment, achieving strong performance with as few as 680 labeled samples and showing improvements over random sampling. Together, the frameworks provide a principled approach to building trustworthy, interpretable, and data-efficient medical AI that generalizes across modalities such as MRI, X-ray, and ultrasound, with robust statistical support and comprehensive ablations. The key ideas are formalized through the joint loss $\mathcal{L}_{total} = \mathcal{L}_{proto} + \alpha \cdot \mathcal{L}_{exp}$ and the acquisition score $\text{Score}(x) = \lambda \mathcal{H}(x) + (1-\lambda) D_{exp}(x)$, enabling a closed-loop learning process where explainability guides both training and data collection.

Abstract

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.

Figures (7)

• Figure 1: Expert-Guided Explainable Few-Shot Learning (EGxFSL) Framework. The pipeline consists of six key steps: (1) Input processing where support sets ($x_s, y_s$) and query samples ($x_q, y_q$) are fed into the system, with expert annotations available for both support and query sets. (2) Feature extraction using a DenseNet-121 backbone within the prototypical network to generate embeddings in a learned embedding space. (3) Prototypical classification where class prototypes are computed from support embeddings and used to classify query samples; (4) GradCAM attention map generation ($G(x_q)$) for query samples to visualize model focus regions; (5) Explainability alignment where Dice loss is computed between the generated GradCAM attention maps and expert annotations from query samples ($M_q$) to measure alignment; (6) Joint optimization using the combined loss function $\mathcal{L}_{total}$ (Equation \ref{['eq:total_loss']}) to simultaneously train for accurate classification and clinically-aligned attention for trustworthy, explainable diagnosis.
• Figure 2: Explainability-Guided Active Learning (xGAL) Framework. The framework operates in an iterative cycle with five key components: (1) Starting with an unlabeled pool $U=\{x_i\}$ where expert annotations are available for selected samples; (2) A composite scoring function that combines classification uncertainty $\mathcal{H}(x)$ and explanation misalignment $D_{\text{exp}}(x)$; (3) Composite acquisition score $\text{Score}(x)$ where $\lambda$ balances uncertainty and misalignment; (4) Selection of top-K samples with highest composite scores, followed by expert annotation to obtain both class labels and diagnostic ROI masks for support and query sets; (5) Model retraining using the EGxFSL framework (Figure \ref{['fig:exp-guided']}). This iterative system progressively improves both classification accuracy and explanation quality with minimal labeled data.
• Figure 3: Effect of explanation weight parameter $\alpha$ on classification accuracy across three medical imaging datasets using our EGxFSL framework. Results obtained using 3-shot FSL configuration trained for 7 epochs with 60 episodes per epoch. Optimal performance is achieved at $\alpha = 0.10$ across all datasets (BraTS: 83.61%, VinDr-CXR: 73.29%, SIIM-COVID: 55.83%), demonstrating that moderate explanation supervision effectively balances classification accuracy with attention alignment. Both under-supervision ($\alpha < 0.10$) and over-supervision ($\alpha > 0.10$) lead to performance degradation, confirming the importance of proper weighting between prototypical loss and explanation alignment loss in Equation \ref{['eq:total_loss']}.
• Figure 4: Comparison of Grad-CAM heatmaps on BraTS, VinDr-CXR, SIIM-COVID datasets. From left to right in each sample: Input Image, Expert Annotation, Guided (EGxFSL) Model Heatmap, Non-Guided Model Heatmap. The guided model consistently focuses on clinically relevant regions as annotated by experts, while the non-guided model highlights irrelevant areas.
• Figure 5: Impact of balancing parameter $\lambda$ on xGAL performance across three medical imaging datasets. The parameter controls the trade-off between classification uncertainty and explanation misalignment in sample acquisition (Equation \ref{['eq:acc_score']}). Optimal performance is achieved at $\lambda = 0.5$ for BraTS and SIIM-COVID, and $\lambda = 0.6$ for VinDr-CXR, demonstrating that balanced consideration of both uncertainty and interpretability yields superior sample selection compared to either criterion alone.