Exploring Entropy-based Active Learning for Fair Brain Segmentation

Technical Summary: Exploring Entropy-based Active Learning for Fair Brain Segmentation

Problem Statement
Active learning (AL) is a critical strategy for mitigating the high costs of expert annotation in medical image segmentation. However, standard uncertainty-based AL methods typically optimize for global performance metrics, often ignoring disparities in performance across groups defined by sensitive attributes (e.g., race, sex). While fair active learning has been explored in classification tasks, its application to medical image segmentation remains largely unaddressed. Existing segmentation models often exhibit demographic disparities, where performance degrades for underrepresented groups or those with specific anatomical variations. Furthermore, naive uncertainty sampling in segmentation tends to select redundant images or outliers, and standard entropy calculations can be biased by anatomical volume differences rather than true model uncertainty.

Methodology
The authors propose a Fairness-aware Active Learning Framework specifically designed for brain MRI segmentation. The core of the method is a Weighted Localized Entropy selection strategy that modulates voxel-wise uncertainty based on current group-specific performance.

1. Data Generation: To control for bias, the study utilizes synthetic T1-weighted brain MRIs generated via the SimBA framework. Two groups are defined:

   • Group 1: Cases with global deformations plus a localized deformation in the left caudate (simulating a "bias" or specific anatomical variation).
   • Group 2: Cases with global deformations only.
     Experiments were conducted under "strong bias" ($\mu=4$) and "weak bias" ($\mu=2$) conditions.

2. Weighted Localized Entropy Strategy:

   • Localized Entropy: To decouple true epistemic uncertainty from anatomical volume variances, entropy is computed only within a dilated Region of Interest (ROI) mask around the target structure (left caudate). The entropy is averaged over the ROI and normalized by region size ($\hat{H}$), preventing larger structures from dominating the selection score.
   • Group-Aware Weighting: The framework calculates a standardized performance score ($z_g$) for each group $g$ based on the Dice Similarity Coefficient (DSC) on the current labeled set. Groups with lower performance (lower DSC) receive higher weights ($w_g$) via a softmax transformation.
   • Acquisition Score: The final score for selecting a sample $x_g$ is the product of the group weight and the localized entropy: $score(x_g) = w_g \cdot \hat{H}$. This prioritizes samples from groups where the model currently underperforms.

3. Experimental Setup: A 3D U-Net was trained using an AL loop starting from small labeled sets (10 samples) with varying initial group compositions (balanced, 80/20, 20/80). The framework was compared against Random Sampling (RS), Mean Entropy Sampling, and Localized Entropy Sampling.

Key Contributions

• First Fair AL for Segmentation: This work addresses the intersection of fair active learning and medical image segmentation, a domain previously unexplored compared to classification.
• Novel Selection Strategy: Introduction of a Weighted Localized Entropy strategy that dynamically re-weights uncertainty based on group-specific DSC performance, ensuring the AL process focuses on underperforming subgroups.
• Decoupling Uncertainty from Anatomy: The use of masked, scaled entropy within a dilated ROI to ensure selection is driven by boundary uncertainty rather than structural volume differences.

Results
Experiments on synthetic datasets demonstrated that the proposed method significantly reduces performance disparities between groups compared to random sampling and standard uncertainty methods.

• Equity-Scaled Performance: The Weighted Localized Entropy strategy consistently achieved the highest Equity-Scaled Segmentation Performance (ESSP) across all initialization regimes and bias strengths.
• Disparity Reduction: Relative to standard entropy sampling, the proposed method reduced the disparity metric ($\Delta$) by 75% in the strong bias setting and 86% in the weak bias setting at the final labeling budget.
• Selection Dynamics: The method successfully prioritized samples from the underperforming Group 1 (the morphology-challenged subgroup), whereas global entropy often selected more Group 2 samples, and random sampling maintained a 50/50 balance.
• Baseline Findings: Surprisingly, baseline experiments showed that training exclusively on the more challenging Group 1 data yielded better ESSP than balanced training, suggesting that over-representing the difficult morphology acts as a "fairness anchor" without severely compromising overall utility.

Significance and Claims
The authors claim this work offers an efficient strategy to train more equitable models in resource-constrained environments where annotation budgets are limited. The framework is notable for being lightweight; it relies solely on labeled-set performance to estimate group weights, requiring no additional fairness classifiers or expensive representativeness modeling in feature spaces.

The paper modestly acknowledges limitations, primarily the use of synthetic data to explicitly control bias magnitude. The authors note that real-world medical data involves complex, hard-to-disentangle biases (e.g., scanner artifacts). They argue that while their results support the utility of weighted sampling to avoid group-level performance bias, translating this to real-world scenarios requires identifying sensitive attributes based on prior hypotheses and reliable reference segmentations. The study concludes that adaptive selection pressure toward underperforming groups is a viable path toward fair medical image segmentation.