Understanding-informed Bias Mitigation for Fair CMR Segmentation

TL;DR

This study addresses ethnicity bias in AI-based cine CMR segmentation by evaluating multiple bias-mitigation strategies (oversampling, reweighing, Group DRO) and introducing root-cause–informed cropping to remove non-heart features driving bias. It demonstrates that oversampling effectively reduces group disparities and that cropping, especially a cascaded approach, boosts overall accuracy while diminishing bias; combining cropping with oversampling yields further gains. The work includes extensive internal and external validation, showing strong segmentation performance and minimal bias on an external clinical dataset, and argues that fairness-accuracy trade-offs can be avoided with a bias-understanding approach. The findings have practical implications for translating bias-mitigation strategies to clinical CMR analysis, enabling more equitable biomarker assessment and treatment planning.

Abstract

Artificial intelligence (AI) is increasingly being used for medical imaging tasks. However, there can be biases in AI models, particularly when they are trained using imbalanced training datasets. One such example has been the strong ethnicity bias effect in cardiac magnetic resonance (CMR) image segmentation models. Although this phenomenon has been reported in a number of publications, little is known about the effectiveness of bias mitigation algorithms in this domain. We aim to investigate the impact of common bias mitigation methods to address bias between Black and White subjects in AI-based CMR segmentation models. Specifically, we use oversampling, importance reweighing and Group DRO as well as combinations of these techniques to mitigate the ethnicity bias. Second, motivated by recent findings on the root causes of AI-based CMR segmentation bias, we evaluate the same methods using models trained and evaluated on cropped CMR images. We find that bias can be mitigated using oversampling, significantly improving performance for the underrepresented Black subjects whilst not significantly reducing the majority White subjects' performance. Using cropped images increases performance for both ethnicities and reduces the bias, whilst adding oversampling as a bias mitigation technique with cropped images reduces the bias further. When testing the models on an external clinical validation set, we find high segmentation performance and no statistically significant bias.

Figures (5)

• Figure 1: 'Cascaded' approach to bias mitigation based on using cropped images. Images (a) are first segmented using a full-image nnU-Net (b) to produce a segmentation (c). This segmentation is then used to crop the images (d) and used in another nnU-Net model trained using ground truth segmentation-based cropped images (e) to produce the final segmentations (f).
• Figure 2: Overall DSC for bias mitigation methods on uncropped images. The dashed line indicates median DSC for White test subjects.
• Figure S1: The effect of changing the number of Black subjects in the training dataset.
• Figure S2: The effect of changing level of oversampling of Black subjects in the training dataset.
• Figure S3: Median DSC against fairness gap for different levels of oversampling in experiments using oversampling and cascaded cropping + oversampling. The fairness gap is calculate by subtracting the median DSC for Black subjects from the median DSC for White subjects.