Understanding-informed Bias Mitigation for Fair CMR Segmentation

Technical Summary: Understanding-informed Bias Mitigation for Fair CMR Segmentation

Problem Statement
Artificial intelligence (AI) models for medical imaging, specifically cardiac magnetic resonance (CMR) segmentation, frequently exhibit bias against protected attributes such as ethnicity. Previous research has established that models trained on imbalanced datasets often underperform for minority groups (e.g., Black subjects compared to White subjects), leading to disparate clinical outcomes. While generic bias mitigation algorithms exist, they often suffer from a "fairness-accuracy trade-off," where improving fairness for minority groups degrades overall performance or the performance of the majority group. Furthermore, these "blind" mitigation approaches often fail to generalize to external validation sets because they do not address the underlying root causes of the bias. Recent investigations suggest that distributional differences in image areas outside the heart contribute significantly to biased representations in CMR segmentation models, yet this insight has not been fully leveraged for mitigation.

Methodology
The authors conducted a comprehensive evaluation of bias mitigation strategies using a dataset of 5,778 subjects from the UK Biobank, curated to include a significant imbalance between Black (15 subjects) and White (4,221 subjects) training samples. An external validation set of 1542 subjects from St. Thomas' Hospital was used to test generalizability.

The study evaluated the following approaches using a 2D nnU-Net architecture:

1. Generic Mitigation Techniques:
   • Oversampling: Increasing the sampling frequency of the minority (Black) group during batch selection to equalize group sizes.
   • Importance Reweighing: Applying inverse-proportional weights to the loss function based on group size.
   • Group Distributionally Robust Optimization (Group DRO): Optimizing the loss for the worst-performing group.
   • Combinations: Pairing the above methods (e.g., Oversampling + Group DRO).
2. Understanding-Informed Mitigation (Root Cause Addressing):
   • Cropping: Based on prior findings that non-cardiac regions drive bias, the authors trained models on images cropped to the heart region.
   • Ground-Truth Cropping: Using manual segmentation masks to define crop boundaries (serving as an upper-bound performance benchmark).
   • Cascaded Cropping: A deployable approach where an initial nnU-Net segmenter on full images predicts a bounding box, which is then used to crop the image for a second, specialized nnU-Net trained on cropped data.

Performance was measured using the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD). Fairness was quantified using the Fairness Gap (FG) and Skewed Error Ratio (SER).

Key Results

• Generic Methods: On internal validation, oversampling was the only generic method that significantly improved performance for Black subjects without statistically reducing performance for White subjects, effectively eliminating the fairness gap. Reweighing degraded performance for both groups, and Group DRO improved Black subject performance but not enough to match White subjects. Combinations of methods generally did not outperform oversampling alone.
• Cropping Impact: Training on cropped images (removing non-cardiac context) significantly improved performance for both ethnicities and reduced bias compared to full-image training.
• Cascaded Approach: The "cascaded" cropping method (using a predicted bounding box) achieved performance comparable to, and in some cases better than, the ground-truth cropping upper bound. This suggests the intermediate segmentation provided a larger, more informative crop than the manual ground truth.
• External Validation: On the external clinical dataset, the baseline model showed no statistically significant bias, and oversampling did not further improve fairness or performance, highlighting the difficulty of generic mitigation under domain shift. However, the cascaded cropping approach with oversampling maintained high performance for both groups and further reduced the fairness gap, outperforming the uncropped baseline.

Key Contributions

1. Comprehensive Evaluation: The paper presents the most extensive investigation to date of multiple bias mitigation methods and their combinations specifically for CMR segmentation.
2. Root-Cause-Informed Mitigation: It is one of the first studies to apply knowledge of the bias's root cause (non-cardiac distributional shifts) to develop a mitigation strategy (cropping), demonstrating that understanding the bias mechanism is crucial for effective correction.
3. Validation of Trade-off Avoidance: The study demonstrates that by using an understanding-informed approach (cascaded cropping), the typical "fairness-accuracy trade-off" can be avoided. The proposed method improves fairness without sacrificing, and in some cases enhancing, overall segmentation accuracy.
4. Clinical Scalability: The authors propose and validate a "cascaded" cropping pipeline that does not require ground-truth masks at inference time, making the bias mitigation strategy deployable in real-world clinical workflows.

Significance and Claims
The authors claim that this work challenges the prevailing notion that bias mitigation in medical AI inevitably requires a trade-off between fairness and accuracy. They argue that generic algorithms often fail because they treat bias as a statistical artifact rather than a structural issue rooted in the data distribution. By identifying that non-cardiac features drive the ethnicity bias in CMR, the authors show that removing these features (via cropping) resolves the bias while improving model robustness.

The paper concludes that for CMR segmentation, a "bias understanding-informed" approach is superior to generic mitigation. This finding is presented as a motivation for future research to prioritize understanding the specific causes of bias in individual medical imaging scenarios rather than relying solely on generic algorithmic fixes. The work suggests that such targeted approaches can lead to robust, fair, and clinically viable AI systems without compromising diagnostic performance.