Source-linked AI summary
Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation
Esther Puyol-Anton, Bram Ruijsink, Stefan K. Piechnik, Stefan Neubauer, Steffen E. Petersen, Reza Razavi, Andrew P. King
TL;DR
Fairness of cardiac MR segmentation across demographic groups had not been studied, despite the clinical relevance of cardiac imaging. The authors compare five approaches to assess and mitigate race- and gender-related bias. They report racial but not gender bias in the UK Biobank setting, with improved fairness from all three mitigation strategies.
Problem
Although cardiac structure and function vary with race and gender, fairness in AI-based cardiac image analysis had not been investigated.
Source: §1 · p. 2Method
The study compares a fairness-through-unawareness baseline and balanced database with three mitigation strategies, including stratified sampling, joint protected-attribute classification, and group-specific models.
Source: §3 · pp. 4–5Results
Takeaways & Limitations
Manual clinician correction is time-consuming and error-prone, so it does not eliminate bias completely.
Source: §5 · p. 9Protected-group models require protected attributes at inference; where those attributes are unavailable, the authors identify stratified sampling or fair meta-learning as alternatives.
Source: §5 · p. 9
Abstract
from arXiv · showhide
The subject of "fairness" in artificial intelligence (AI) refers to assessing AI algorithms for potential bias based on demographic characteristics such as race and gender, and the development of algorithms to address this bias. Most applications to date have been in computer vision, although some work in healthcare has started to emerge. The use of deep learning (DL) in cardiac MR segmentation has led to impressive results in recent years, and such techniques are starting to be translated into clinical practice. However, no work has yet investigated the fairness of such models. In this work, we perform such an analysis for racial/gender groups, focusing on the problem of training data imbalance, using a nnU-Net model trained and evaluated on cine short axis cardiac MR data from the UK Biobank dataset, consisting of 5,903 subjects from 6 different racial groups. We find statistically significant differences in Dice performance between different racial groups. To reduce the racial bias, we investigated three strategies: (1) stratified batch sampling, in which batch sampling is stratified to ensure balance between racial groups; (2) fair meta-learning for segmentation, in which a DL classifier is trained to classify race and jointly optimized with the segmentation model; and (3) protected group models, in which a different segmentation model is trained for each racial group. We also compared the results to the scenario where we have a perfectly balanced database. To assess fairness we used the standard deviation (SD) and skewed error ratio (SER) of the average Dice values. Our results demonstrate that the racial bias results from the use of imbalanced training data, and that all proposed bias mitigation strategies improved fairness, with the best SD and SER resulting from the use of protected group models.
1 Introduction
Deep learning cardiac MR segmentation is moving toward clinical use, but its fairness across demographic groups had not been investigated. This study examines racial and gender bias and proposes mitigation strategies.
1 Introduction
- Cardiac MR segmentation models have reported human-level performance and are beginning to move toward wider clinical translation.Proposed uses include automated cardiac functional quantification.Source: §1 · p. 2
- Cardiac structure, function, and cardiovascular disease vary by race and gender, while disparities in disease detection may contribute to these differences.The cited MESA study found racial differences in regional left ventricular systolic function.Source: §1 · p. 2
- Before this study, fairness in AI-based cardiac image analysis had not been investigated, despite growing fairness research in AI.The authors identify gender-related radiomics analysis as the closest related work.Source: §1 · p. 2
- The study analyzes race and gender imbalance in cardiac MR segmentation and proposes strategies to mitigate observed racial performance differences.The authors describe it as the first fairness analysis of DL-based cardiac MR segmentation.Source: §1 · p. 2
2 Background
Fairness methods either use protected attributes or omit them, and bias mitigation can intervene in data preparation, model training, or predictions. Prior work had not addressed bias mitigation for image segmentation.
2 Background
- Prior fairness research largely focused on classification; Hwang et al. [7] studied image-to-image translation, but the authors found no prior image-segmentation bias mitigation work.Source: §2 · p. 2
- Fairness through awareness uses protected attributes to make predictions independently by group, whereas fairness through unawareness assumes omitting those attributes will yield fair decisions.The authors note that unawareness can fail when protected attributes correlate with other data variables.Source: §2 · p. 3
- Bias mitigation can intervene through preprocessing the training data, in-processing model training, or post-processing output predictions.Preprocessing includes balancing or reweighting data; in-processing modifies learning, and post-processing adjusts outputs.Source: §2 · p. 3
- A protected attribute is one for which fairness is sought, such as race or gender; a protected group shares the same attribute value(s).Source: §2 · p. 3
3 Methods
The study compares a fairness-through-unawareness baseline and a smaller balanced database with three awareness-based mitigation approaches. It evaluates segmentation overlap and differences in performance across protected groups.
3 Methods
- Five approaches compare an unawareness baseline and balanced-database test with stratified sampling, fair meta-learning, and protected-group models.The balanced database tests whether observed bias is due to training-data imbalance.Source: §3 · p. 4, Figure 1 · p. 4
- The baseline trains nnU-Net without race or gender attributes; the balanced comparison trains on 522 subjects, selecting 87 from each racial group.Source: §3 · pp. 4–5
- Stratified batch sampling selects training samples so each protected group is equally represented in every batch.Source: §3 · p. 5
- Fair meta-learning jointly trains segmentation and protected-attribute classification networks, while protected-group models fine-tune separate segmentation models for each group.The protected-group approach assumes protected attributes are available during training and inference.Source: §3 · p. 5
- Segmentation overlap is measured by average Dice similarity across LVBP, LVM, and RVBP; fairness is assessed using SD and SER across protected groups.SD reflects dispersion in group-average Dice, while SER is the ratio of the highest to lowest group error rate.Source: §3 · pp. 5–6
4 Materials and Experiments
In UK Biobank cardiac MR data, baseline segmentation performance differed by race but not gender. The experiments compared bias-mitigation strategies, with protected-group models yielding the lowest SD and SER among the mitigation approaches.
4 Materials and Experiments
- The study used ED and ES short-axis cine cardiac MR data from 5,903 UK Biobank subjects; Figure 2 summarizes the dataset’s gender and racial percentages.For approaches other than the balanced-database comparison, the train/validation/test split was 4,723/590/590 subjects.Source: §4 · p. 6, Figure 2 · p. 6
- 93.51% average DSC for white subjects versus less than 86% for black and less than 85% for mixed-race subjects; baseline showed no significant gender bias.The comparison uses average DSC; Table 1 reports full-population and protected-group results.Source: §4 · pp. 6–7, Table 1 · p. 7
- The smaller balanced-database comparison supported the link between group representation and performance, but significantly reduced accuracy for every racial group.The balanced comparison used a smaller training database than the other approaches.Source: §4 · pp. 6–7
- All three mitigation strategies reduced racial bias; protected-group models had the lowest SD and SER, while stratified batch sampling was best without protected-attribute knowledge at inference.Protected-group models require protected attributes at inference time.Source: §4 · p. 7, Table 2 · p. 8
- Figure 3 shows end-diastolic segmentation examples for each racial group, with high-DSC examples above low-DSC examples.These are illustrative samples rather than a quantitative group comparison.Source: §4 · p. 8, Figure 3 · p. 8
5 Discussion
The discussion reports racial but not gender bias in UK Biobank cardiac MR segmentation and very low bias with balanced training data. It describes three strategies that improved fairness, while noting that the best-performing strategy requires protected-attribute information at inference.
5 Discussion
- UK Biobank models showed racial but not gender bias, while a balanced database showed very low bias, supporting training-data imbalance as a source of racial bias.The authors note that clinician modification is time-consuming and error-prone, so it does not eliminate bias completely.Source: §5 · pp. 8–9
- Because clinician oversight may decrease or disappear as models gain clinical acceptance, reducing bias may become critical.Current clinician modification is time-consuming and prone to error, and therefore does not eliminate bias completely.Source: §5 · p. 9
- All three proposed strategies improved fairness, with lower SD and SER than the baseline: stratified batches, fair meta-learning, and separate protected-group models.The strategies balance group representation within batches, jointly train a protected-attribute classifier, or train independent segmentation models for each group.Source: §5 · p. 9
- The best-performing protected-group strategy requires protected-attribute information at inference; the other two strategies remain options when it is unavailable.The authors describe Approach 3 as a possible “best achievable” model when protected attributes are inadequately registered.Source: §5 · p. 9
- The authors identify combinations of pre-processing, in-processing, and post-processing strategies and other imaging modalities as future directions.The three proposed techniques are described as novel in image segmentation, though inspired by fairness work in classification.Source: §5 · p. 9