Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI

Lucas Fidon; Michael Aertsen; Nada Mufti; Thomas Deprest; Doaa Emam; Frédéric Guffens; Ernst Schwartz; Michael Ebner; Daniela Prayer; Gregor Kasprian; Anna L. David; Andrew Melbourne; Sébastien Ourselin; Jan Deprest; Georg Langs; Tom Vercauteren

doi:10.1007/978-3-030-87735-4_25

Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI

Lucas Fidon^*, Michael Aertsen, Nada Mufti, Thomas Deprest, Doaa Emam, Frédéric Guffens, Ernst Schwartz, Michael Ebner, Daniela Prayer, Gregor Kasprian, Anna L. David, Andrew Melbourne, Sébastien Ourselin, Jan Deprest, Georg Langs, Tom Vercauteren

^*Corresponding author for this work

Research output: Contribution to journal › Article › peer-review

13 Citations (Scopus)

Abstract

The performance of deep neural networks typically increases with the number of training images. However, not all images have the same importance towards improved performance and robustness. In fetal brain MRI, abnormalities exacerbate the variability of the developing brain anatomy compared to non-pathological cases. A small number of abnormal cases, as is typically available in clinical datasets used for training, are unlikely to fairly represent the rich variability of abnormal developing brains. This leads machine learning systems trained by maximizing the average performance to be biased toward non-pathological cases. This problem was recently referred to as hidden stratification. To be suited for clinical use, automatic segmentation methods need to reliably achieve high-quality segmentation outcomes also for pathological cases. In this paper, we show that the state-of-the-art deep learning pipeline nnU-Net has difficulties to generalize to unseen abnormal cases. To mitigate this problem, we propose to train a deep neural network to minimize a percentile of the distribution of per-volume loss over the dataset. We show that this can be achieved by using Distributionally Robust Optimization (DRO). DRO automatically reweights the training samples with lower performance, encouraging nnU-Net to perform more consistently on all cases. We validated our approach using a dataset of 368 fetal brain T2w MRIs, including 124 MRIs of open spina bifida cases and 51 MRIs of cases with other severe abnormalities of brain development.

Original language	English
Pages (from-to)	263-273
Number of pages	11
Journal	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
DOIs	https://doi.org/10.1007/978-3-030-87735-4_25
Publication status	Published - 2021
Event	3rd International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, UNSURE 2021, and the 6th International Workshop on Perinatal, Preterm and Paediatric Image Analysis, PIPPI 2021, held in conjunction with the 24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021 - Virtual, Online Duration: 1 Oct 2021 → 1 Oct 2021

Access to Document

10.1007/978-3-030-87735-4_25

Cite this

Fidon, L., Aertsen, M., Mufti, N., Deprest, T., Emam, D., Guffens, F., Schwartz, E., Ebner, M., Prayer, D., Kasprian, G., David, A. L., Melbourne, A., Ourselin, S., Deprest, J., Langs, G., & Vercauteren, T. (2021). Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 263-273. https://doi.org/10.1007/978-3-030-87735-4_25

@article{f56759bb40074163931667551d0fd7cf,

title = "Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI",

abstract = "The performance of deep neural networks typically increases with the number of training images. However, not all images have the same importance towards improved performance and robustness. In fetal brain MRI, abnormalities exacerbate the variability of the developing brain anatomy compared to non-pathological cases. A small number of abnormal cases, as is typically available in clinical datasets used for training, are unlikely to fairly represent the rich variability of abnormal developing brains. This leads machine learning systems trained by maximizing the average performance to be biased toward non-pathological cases. This problem was recently referred to as hidden stratification. To be suited for clinical use, automatic segmentation methods need to reliably achieve high-quality segmentation outcomes also for pathological cases. In this paper, we show that the state-of-the-art deep learning pipeline nnU-Net has difficulties to generalize to unseen abnormal cases. To mitigate this problem, we propose to train a deep neural network to minimize a percentile of the distribution of per-volume loss over the dataset. We show that this can be achieved by using Distributionally Robust Optimization (DRO). DRO automatically reweights the training samples with lower performance, encouraging nnU-Net to perform more consistently on all cases. We validated our approach using a dataset of 368 fetal brain T2w MRIs, including 124 MRIs of open spina bifida cases and 51 MRIs of cases with other severe abnormalities of brain development.",

author = "Lucas Fidon and Michael Aertsen and Nada Mufti and Thomas Deprest and Doaa Emam and Fr{\'e}d{\'e}ric Guffens and Ernst Schwartz and Michael Ebner and Daniela Prayer and Gregor Kasprian and David, {Anna L.} and Andrew Melbourne and S{\'e}bastien Ourselin and Jan Deprest and Georg Langs and Tom Vercauteren",

note = "Funding Information: This project has received funding from the European Union?s Horizon 2020 research and innovation program under the Marie Sk?lodowska-Curie grant agreement TRABIT No 765148. This work was supported by core and project funding from the Wellcome [203148/Z/16/Z; 203145Z/16/Z; WT101957], and EPSRC [NS/A000049/1; NS/A000050/1; NS/A000027/1]. TV is supported by a Medtronic/RAEng Research Chair [RCSRF1819\7\34]. Publisher Copyright: {\textcopyright} 2021, Springer Nature Switzerland AG.; 3rd International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, UNSURE 2021, and the 6th International Workshop on Perinatal, Preterm and Paediatric Image Analysis, PIPPI 2021, held in conjunction with the 24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021 ; Conference date: 01-10-2021 Through 01-10-2021",

year = "2021",

doi = "10.1007/978-3-030-87735-4_25",

language = "English",

pages = "263--273",

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

issn = "0302-9743",

publisher = "Springer",

}

Fidon, L, Aertsen, M, Mufti, N, Deprest, T, Emam, D, Guffens, F, Schwartz, E, Ebner, M, Prayer, D, Kasprian, G, David, AL, Melbourne, A , Ourselin, S, Deprest, J, Langs, G & Vercauteren, T 2021, 'Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI', Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), pp. 263-273. https://doi.org/10.1007/978-3-030-87735-4_25

TY - JOUR

T1 - Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI

AU - Fidon, Lucas

AU - Aertsen, Michael

AU - Mufti, Nada

AU - Deprest, Thomas

AU - Emam, Doaa

AU - Guffens, Frédéric

AU - Schwartz, Ernst

AU - Ebner, Michael

AU - Prayer, Daniela

AU - Kasprian, Gregor

AU - David, Anna L.

AU - Melbourne, Andrew

AU - Ourselin, Sébastien

AU - Deprest, Jan

AU - Langs, Georg

AU - Vercauteren, Tom

N1 - Funding Information: This project has received funding from the European Union?s Horizon 2020 research and innovation program under the Marie Sk?lodowska-Curie grant agreement TRABIT No 765148. This work was supported by core and project funding from the Wellcome [203148/Z/16/Z; 203145Z/16/Z; WT101957], and EPSRC [NS/A000049/1; NS/A000050/1; NS/A000027/1]. TV is supported by a Medtronic/RAEng Research Chair [RCSRF1819\7\34]. Publisher Copyright: © 2021, Springer Nature Switzerland AG.

PY - 2021

Y1 - 2021

N2 - The performance of deep neural networks typically increases with the number of training images. However, not all images have the same importance towards improved performance and robustness. In fetal brain MRI, abnormalities exacerbate the variability of the developing brain anatomy compared to non-pathological cases. A small number of abnormal cases, as is typically available in clinical datasets used for training, are unlikely to fairly represent the rich variability of abnormal developing brains. This leads machine learning systems trained by maximizing the average performance to be biased toward non-pathological cases. This problem was recently referred to as hidden stratification. To be suited for clinical use, automatic segmentation methods need to reliably achieve high-quality segmentation outcomes also for pathological cases. In this paper, we show that the state-of-the-art deep learning pipeline nnU-Net has difficulties to generalize to unseen abnormal cases. To mitigate this problem, we propose to train a deep neural network to minimize a percentile of the distribution of per-volume loss over the dataset. We show that this can be achieved by using Distributionally Robust Optimization (DRO). DRO automatically reweights the training samples with lower performance, encouraging nnU-Net to perform more consistently on all cases. We validated our approach using a dataset of 368 fetal brain T2w MRIs, including 124 MRIs of open spina bifida cases and 51 MRIs of cases with other severe abnormalities of brain development.

AB - The performance of deep neural networks typically increases with the number of training images. However, not all images have the same importance towards improved performance and robustness. In fetal brain MRI, abnormalities exacerbate the variability of the developing brain anatomy compared to non-pathological cases. A small number of abnormal cases, as is typically available in clinical datasets used for training, are unlikely to fairly represent the rich variability of abnormal developing brains. This leads machine learning systems trained by maximizing the average performance to be biased toward non-pathological cases. This problem was recently referred to as hidden stratification. To be suited for clinical use, automatic segmentation methods need to reliably achieve high-quality segmentation outcomes also for pathological cases. In this paper, we show that the state-of-the-art deep learning pipeline nnU-Net has difficulties to generalize to unseen abnormal cases. To mitigate this problem, we propose to train a deep neural network to minimize a percentile of the distribution of per-volume loss over the dataset. We show that this can be achieved by using Distributionally Robust Optimization (DRO). DRO automatically reweights the training samples with lower performance, encouraging nnU-Net to perform more consistently on all cases. We validated our approach using a dataset of 368 fetal brain T2w MRIs, including 124 MRIs of open spina bifida cases and 51 MRIs of cases with other severe abnormalities of brain development.

UR - http://www.scopus.com/inward/record.url?scp=85117074289&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-87735-4_25

DO - 10.1007/978-3-030-87735-4_25

M3 - Article

AN - SCOPUS:85117074289

SN - 0302-9743

SP - 263

EP - 273

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)

T2 - 3rd International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, UNSURE 2021, and the 6th International Workshop on Perinatal, Preterm and Paediatric Image Analysis, PIPPI 2021, held in conjunction with the 24th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2021

Y2 - 1 October 2021 through 1 October 2021

ER -

Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI

Abstract

Access to Document

Other files and links

Fingerprint

Cite this