Investigating Racial Bias in AI-Driven Cardiac Imaging
Artificial intelligence increasingly supports segmentation in cardiac magnetic resonance (CMR) imaging, which is crucial for diagnosing cardiovascular conditions. Yet, concerns about racial bias remain due to imbalanced training datasets. Research is uncovering the root causes of this bias to build fairer, more equitable AI models.
Bias and the Limitations of AI Models
AI models reflect the data they learn from. If a model trains mostly on images from one group, it performs better for that group. A key issue is shortcut learning, where models rely on irrelevant features—like race or scanner types—instead of true disease indicators. This leads to misdiagnosis or unequal treatment.
For example, if images from one scanner are linked to severe illness, the AI might associate scanner-specific artifacts with disease severity rather than clinical signs. Similarly, if disease prevalence varies among racial groups in training data, the model might use race as a proxy, causing biased predictions.
Study Overview: Race Classification in CMR Images
The study tested whether AI could classify a patient's race from CMR images. It found that both raw images and segmentations contained enough information for accurate race classification. Surprisingly, the model based decisions mostly on areas outside the heart, such as subcutaneous fat and image artifacts.
When images were cropped to focus only on the heart, race classification accuracy dropped significantly. Blurring the heart entirely did not prevent the model from classifying race accurately, confirming that race-related features largely reside outside cardiac structures.
Can Cropping Images Reduce AI Bias?
Cropping images around the heart lowered race classification accuracy to near chance levels, suggesting that focusing on cardiac structures reduces racial information in the data. However, segmentation models still showed bias despite cropping, indicating underlying factors beyond visible race-related features.
Confounder analysis revealed that variables like MRI scan year and cholesterol levels correlated with segmentation performance for Black patients but not for White patients. This suggests persistent bias despite removing obvious race-related features.
Importantly, increasing Black subjects in training data from 0% to 25% improved model performance for this group without harming accuracy for White patients. This highlights the need for diverse and balanced datasets to achieve fair AI models.
Effective Bias Mitigation Strategies
Oversampling under-represented groups—training the AI more frequently on images from minority populations—helped balance performance across racial groups. Other mitigation techniques were less effective, underscoring the importance of selecting strategies that fit the data and context.
Bias Beyond Cardiac Imaging
Racial and demographic biases extend across medical imaging types. Chest X-ray models demonstrate lower accuracy for younger patients, females, and Black or Hispanic populations, often leading to underdiagnosis. Dermatology AI models perform worse for darker skin tones but improve when fine-tuned with diverse data.
Research also shows that AI can predict race from chest X-rays alone, indicating that “colourblind” training approaches may fail since models detect subtle, non-obvious race features. Because many AI models operate as “black boxes,” increasing transparency and explainability is critical to building trust and identifying biases.
Conclusion
This study shows that racial bias in AI-driven CMR segmentation stems mostly from image differences outside the heart, such as body fat composition. Although cropping images to focus on the heart reduces bias, it does not remove it entirely. Improving dataset diversity and using targeted bias mitigation methods are essential steps forward.
These findings apply beyond cardiac imaging, with similar patterns seen in chest X-rays and dermatology. Building equitable AI tools requires diverse data, explainable models, and collaboration across data science, clinical, and imaging fields to ensure fair patient care.
Your membership also unlocks:
