Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation

Esther Puyol-Antón; Bram Ruijsink; Stefan K. Piechnik; Stefan Neubauer; Steffen E. Petersen; Reza Razavi; Andrew P. King

doi:10.1007/978-3-030-87199-4_39

Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation

Esther Puyol-Antón^*, Bram Ruijsink, Stefan K. Piechnik, Stefan Neubauer, Steffen E. Petersen, Reza Razavi, Andrew P. King

^*Corresponding author for this work

Research output: Contribution to journal › Conference paper › peer-review

68 Citations (Scopus)

186 Downloads (Pure)

Abstract

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.

Original language	English
Pages (from-to)	413-423
Number of pages	11
Journal	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	12903
Early online date	21 Sept 2021
DOIs	https://doi.org/10.1007/978-3-030-87199-4_39
Publication status	E-pub ahead of print - 21 Sept 2021
Event	24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021 - Virtual, Online Duration: 27 Sept 2021 → 1 Oct 2021

Keywords

Cardiac MRI
Fair AI
Inequality
Segmentation

Access to Document

10.1007/978-3-030-87199-4_39

Fairness in Cardiac MR_PUYOL-ANTON_Publishedonline21September2021_GREEN AAMAccepted author manuscript, 883 KB

https://arxiv.org/pdf/2106.12387.pdf

Cite this

Puyol-Antón, E., Ruijsink, B., Piechnik, S. K., Neubauer, S., Petersen, S. E., Razavi, R., & King, A. P. (2021). Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12903, 413-423. Advance online publication. https://doi.org/10.1007/978-3-030-87199-4_39

@article{1de27e5e1c9246088796cc12a5ecf64a,

title = "Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation",

abstract = "The subject of {\textquoteleft}fairness{\textquoteright} 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.",

keywords = "Cardiac MRI, Fair AI, Inequality, Segmentation",

author = "Esther Puyol-Ant{\'o}n and Bram Ruijsink and Piechnik, {Stefan K.} and Stefan Neubauer and Petersen, {Steffen E.} and Reza Razavi and King, {Andrew P.}",

note = "Funding Information: Acknowledgements. This work was supported by the EPSRC (EP/R005516/1 and EP/P001009/1), the Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, King's College London (WT 203148/Z/16/Z) and has been conducted using the UK Biobank Resource under application numbers 17806 and 2964. SEP, SN and SKP acknowledge the BHF for funding the manual analysis to create a cardiac MR imaging reference standard for the UK Biobank imaging resource in 5000 CMR scans (PG/14/89/31194). We also acknowledge the following funding sources: the NIHR Biomedical Research Centre at Barts, EU{\textquoteright}s Horizon 2020 (grant no. 825903, euCanSHare project), the CAP-AI programme funded by the ERDF and Barts Charity, HDR UK, the Oxford NIHR Biomedical Research Centre, the Oxford BHF Centre of Research Excellence and the MRC eMed-Lab Medical Bioinformatics infrastructure (MR/L016311/1). Funding Information: This work was supported by the EPSRC (EP/R005516/1 and EP/P001009/1), the Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, King's College London (WT 203148/Z/16/Z) and has been conducted using the UK Biobank Resource under application numbers 17806 and 2964. SEP, SN and SKP acknowledge the BHF for funding the manual analysis to create a cardiac MR imaging reference standard for the UK Biobank imaging resource in 5000 CMR scans (PG/14/89/31194). We also acknowledge the following funding sources: the NIHR Biomedical Research Centre at Barts, EU?s Horizon 2020 (grant no. 825903, euCanSHare project), the CAP-AI programme funded by the ERDF and Barts Charity, HDR UK, the Oxford NIHR Biomedical Research Centre, the Oxford BHF Centre of Research Excellence and the MRC eMed-Lab Medical Bioinformatics infrastructure (MR/L016311/1). Publisher Copyright: {\textcopyright} 2021, Springer Nature Switzerland AG.; 24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021 ; Conference date: 27-09-2021 Through 01-10-2021",

year = "2021",

month = sep,

day = "21",

doi = "10.1007/978-3-030-87199-4_39",

language = "English",

volume = "12903",

pages = "413--423",

journal = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

issn = "0302-9743",

publisher = "Springer",

}

Puyol-Antón, E, Ruijsink, B, Piechnik, SK, Neubauer, S, Petersen, SE, Razavi, R & King, AP 2021, 'Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation', Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 12903, pp. 413-423. https://doi.org/10.1007/978-3-030-87199-4_39

Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation. / Puyol-Antón, Esther; Ruijsink, Bram; Piechnik, Stefan K. et al.
In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 12903, 21.09.2021, p. 413-423.

Research output: Contribution to journal › Conference paper › peer-review

TY - JOUR

T1 - Fairness in Cardiac MR Image Analysis

T2 - 24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021

AU - Puyol-Antón, Esther

AU - Ruijsink, Bram

AU - Piechnik, Stefan K.

AU - Neubauer, Stefan

AU - Petersen, Steffen E.

AU - Razavi, Reza

AU - King, Andrew P.

N1 - Funding Information: Acknowledgements. This work was supported by the EPSRC (EP/R005516/1 and EP/P001009/1), the Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, King's College London (WT 203148/Z/16/Z) and has been conducted using the UK Biobank Resource under application numbers 17806 and 2964. SEP, SN and SKP acknowledge the BHF for funding the manual analysis to create a cardiac MR imaging reference standard for the UK Biobank imaging resource in 5000 CMR scans (PG/14/89/31194). We also acknowledge the following funding sources: the NIHR Biomedical Research Centre at Barts, EU’s Horizon 2020 (grant no. 825903, euCanSHare project), the CAP-AI programme funded by the ERDF and Barts Charity, HDR UK, the Oxford NIHR Biomedical Research Centre, the Oxford BHF Centre of Research Excellence and the MRC eMed-Lab Medical Bioinformatics infrastructure (MR/L016311/1). Funding Information: This work was supported by the EPSRC (EP/R005516/1 and EP/P001009/1), the Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, King's College London (WT 203148/Z/16/Z) and has been conducted using the UK Biobank Resource under application numbers 17806 and 2964. SEP, SN and SKP acknowledge the BHF for funding the manual analysis to create a cardiac MR imaging reference standard for the UK Biobank imaging resource in 5000 CMR scans (PG/14/89/31194). We also acknowledge the following funding sources: the NIHR Biomedical Research Centre at Barts, EU?s Horizon 2020 (grant no. 825903, euCanSHare project), the CAP-AI programme funded by the ERDF and Barts Charity, HDR UK, the Oxford NIHR Biomedical Research Centre, the Oxford BHF Centre of Research Excellence and the MRC eMed-Lab Medical Bioinformatics infrastructure (MR/L016311/1). Publisher Copyright: © 2021, Springer Nature Switzerland AG.

PY - 2021/9/21

Y1 - 2021/9/21

N2 - 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.

AB - 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.

KW - Cardiac MRI

KW - Fair AI

KW - Inequality

KW - Segmentation

UR - http://www.scopus.com/inward/record.url?scp=85116476281&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-87199-4_39

DO - 10.1007/978-3-030-87199-4_39

M3 - Conference paper

AN - SCOPUS:85116476281

SN - 0302-9743

VL - 12903

SP - 413

EP - 423

JO - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

JF - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Y2 - 27 September 2021 through 1 October 2021

ER -

Puyol-Antón E, Ruijsink B, Piechnik SK, Neubauer S, Petersen SE, Razavi R et al. Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2021 Sept 21;12903:413-423. Epub 2021 Sept 21. doi: 10.1007/978-3-030-87199-4_39

Fairness in Cardiac MR Image Analysis: An Investigation of Bias Due to Data Imbalance in Deep Learning Based Segmentation

Abstract

Keywords

Access to Document

Other files and links

Fingerprint

Cite this