Fairness in AI: are deep learning-based CMR segmentation algorithms biased?

Background/Introduction

Artificial intelligence (AI) is providing opportunities to transform cardiovascular medicine. A particular challenge in the application of AI technology is their potential for intrinsic and extrinsic biases, such as those based on gender and/or ethnicity. Unless satisfactorily addressed, these biases could lead to inequalities in early diagnosis, treatments and outcomes. Fairness in AI is a relatively new but fast-growing research field which deals with assessing and addressing potential bias in AI models.

Purpose

To perform the first analysis that assesses bias in AI-based cardiac MR segmentation models in a large-scale database.

Methods

A state-of-the-art deep learning (DL) based segmentation network, the “nnU-Net” framework [1], was used for automatic segmentation of both ventricles and the myocardium from cine short-axis cardiac MR over the full cardiac cycle. The dataset used consisted of end-diastole and end-systole short-axis cine cardiac MR images of 5,903 subjects (61.5±7.1 years). The nnU-Net network was trained and evaluated using a 5-fold cross validation (splits: train 60% / validation 20% / test 20%). Data on race and gender were obtained from the UK Biobank database and their distribution is summarized in Figure 1. To assess gender and racial bias in the segmentation network, we compared the Dice scores - which measure the overlap between manual and automatic segmentations – and the absolute error in measurements of biventricular volumes and function between patients grouped by ethnicity and gender.

Results

Figure 2 shows the Dice scores and the volumetric and functional measures for the full database, stratified by gender and by ethnicity. Results on the overall population showed an excellent agreement between the manual and automatic segmentations which is consistent with previous reported results [2–3]. However, we find statistically significant differences in Dice scores as well as volumetric measures between different ethnicities, showing that the segmentation network is biased against minority racial groups. No significant differences were found in Dice scores between genders. Similarly, for the end diastolic, end systolic volumes and ejection fraction, there were statistically significant differences in absolute error between the overall population and all racial groups except white.

Conclusion(s)

We have shown, for the first time, that racial bias exists in DL-based cardiac MR segmentation models. Our hypothesis is that this bias is a result of the unbalanced nature of the training data, and this is supported by the results which show that there is racial bias but not gender bias when trained using the UK Biobank database, which is gender-balanced but not race-balanced. In this work we want to highlight the potential issue of bias in DL-based image segmentation models when translating into a clinical environment.

Funding Acknowledgement

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): - EPSRC- Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, King's College London Figure 1Figure 2

Large-scale, multi-vendor, multi-protocol, quality-controlled analysis of clinical cine CMR using artificial intelligence

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Advancing Impact Award scheme of the EPSRC Impact Acceleration Account at King’s College London

Background

Artificial intelligence (AI) has the potential to facilitate the automation of CMR analysis for biomarker extraction. However, most AI algorithms are trained on a specific input domain (e.g., scanner vendor or hospital-tailored imaging protocol) and lack the robustness to perform optimally when applied to CMR data from other input domains.

Purpose

To develop and validate a robust CMR analysis tool for automatic segmentation and cardiac function analysis which achieves state-of-the-art performance for multi-vendor short-axis cine CMR images. 

Methods

The current work is an extension of our previously published quality-controlled AI-based tool for cine CMR analysis [1]. We deployed an AI algorithm that is equipped to handle different image sizes and domains automatically - the ‘nnU-Net’ framework [2] - and retrained our tool using the UK Biobank (UKBB) cohort population (n = 4,872) and a large database of clinical CMR studies obtained from two NHS hospitals (n = 3,406). The NHS hospital data came from three different scanner types: Siemens Aera 1.5T (n = 1,419), Philips Achieva 1.5T and 3T (n = 1,160), and Philips Ingenia 1.5T (n = 827). The ‘nnU-net’ was used to segment both ventricles and the myocardium. The proposed method was evaluated on randomly selected test sets from UKBB (n = 488) and NHS (n = 331) and on two external publicly available databases of clinical CMRs acquired on Philips, Siemens, General Electric (GE), and Canon CMR scanners – ACDC (n = 100) [3] and M&Ms (n = 321) [4]. We calculated the Dice scores - which measure the overlap between manual and automatic segmentations - and compared manual vs AI-based measures of biventricular volumes and function. 

Results

 Table 1 shows that the Dice scores for the NHS, ACDC, and M&Ms scans are similar to those obtained in the highly controlled, single vendor and single field strength UKBB scans. Although our AI-based tool was only trained on CMR scans from two vendors (Philips and Siemens), it performs similarly in unseen vendors (GE and Canon). Furthermore, it achieves state-of-the-art performance in online segmentation challenges, without being specifically trained on these databases. Table 1 also shows good agreement between manual and automated clinical measures of ejection fraction and ventricular volume and mass.

Conclusions

We show that our proposed AI-based tool, which combines training on a large-scale multi-domain CMR database with a state-of-the-art AI algorithm, allows us to robustly deal with routine clinical data from multiple centres, vendors, and field strengths. This is a fundamental step for the clinical translation of AI algorithms. Moreover, our method yields a range of additional metrics of cardiac function (filling and ejection rates, regional wall motion, and strain) at no extra computational cost.