Recent progress in Medical Artificial Intelligence (AI) has delivered systems that can reach clinical expert level performance. However, such systems tend to demonstrate sub-optimal öut-of-distribution" performance when evaluated in clinical settings different from the training environment. A common mitigation strategy is to develop separate systems for each clinical setting using site-specific data 1. However, this quickly becomes impractical as medical data is time-consuming to acquire and expensive to annotate 2. Thus, the problem of "data-efficient generalization" presents an ongoing difficulty for Medical AI development. Although progress in representation learning shows promise, their benefits have not been rigorously studied, specifically for out-of-distribution settings. To meet these challenges, we present REMEDIS, a unified representation learning strategy to improve robustness and data-efficiency of medical imaging AI. REMEDIS uses a generic combination of large-scale supervised transfer learning with self-supervised learning and requires little task-specific customization. We study a diverse range of medical imaging tasks and simulate three realistic application scenarios using retrospective data. REMEDIS exhibits significantly improved in-distribution performance with up to 11.5\% relative improvement in diagnostic accuracy over a strong supervised baseline. More importantly, our strategy leads to strong data-efficient generalization of medical imaging AI, matching strong supervised baselines using between 1\% to 33\% of retraining data across tasks. These results suggest that REMEDIS can significantly accelerate the life-cycle of medical imaging AI development thereby presenting an important step forward for medical imaging AI to deliver broad impact.
Azizi et al. - 2022 - Robust and Efficient Medical Imaging with Self-Sup.pdf:files/24/Azizi et al. - 2022 - Robust and Efficient Medical Imaging with Self-Sup.pdf:application/pdf
%0 Generic
%1 azizi_robust_2022
%A Azizi, Shekoofeh
%A Culp, Laura
%A Freyberg, Jan
%A Mustafa, Basil
%A Baur, Sebastien
%A Kornblith, Simon
%A Chen, Ting
%A MacWilliams, Patricia
%A Mahdavi, S. Sara
%A Wulczyn, Ellery
%A Babenko, Boris
%A Wilson, Megan
%A Loh, Aaron
%A Chen, Po-Hsuan Cameron
%A Liu, Yuan
%A Bavishi, Pinal
%A McKinney, Scott Mayer
%A Winkens, Jim
%A Roy, Abhijit Guha
%A Beaver, Zach
%A Ryan, Fiona
%A Krogue, Justin
%A Etemadi, Mozziyar
%A Telang, Umesh
%A Liu, Yun
%A Peng, Lily
%A Corrado, Greg S.
%A Webster, Dale R.
%A Fleet, David
%A Hinton, Geoffrey
%A Houlsby, Neil
%A Karthikesalingam, Alan
%A Norouzi, Mohammad
%A Natarajan, Vivek
%D 2022
%I arXiv
%K - Artificial Computer Intelligence, Learning Machine Pattern Recognition, Science Vision and
%T Robust and Efficient Medical Imaging with Self-Supervision
%U http://arxiv.org/abs/2205.09723
%X Recent progress in Medical Artificial Intelligence (AI) has delivered systems that can reach clinical expert level performance. However, such systems tend to demonstrate sub-optimal öut-of-distribution" performance when evaluated in clinical settings different from the training environment. A common mitigation strategy is to develop separate systems for each clinical setting using site-specific data 1. However, this quickly becomes impractical as medical data is time-consuming to acquire and expensive to annotate 2. Thus, the problem of "data-efficient generalization" presents an ongoing difficulty for Medical AI development. Although progress in representation learning shows promise, their benefits have not been rigorously studied, specifically for out-of-distribution settings. To meet these challenges, we present REMEDIS, a unified representation learning strategy to improve robustness and data-efficiency of medical imaging AI. REMEDIS uses a generic combination of large-scale supervised transfer learning with self-supervised learning and requires little task-specific customization. We study a diverse range of medical imaging tasks and simulate three realistic application scenarios using retrospective data. REMEDIS exhibits significantly improved in-distribution performance with up to 11.5\% relative improvement in diagnostic accuracy over a strong supervised baseline. More importantly, our strategy leads to strong data-efficient generalization of medical imaging AI, matching strong supervised baselines using between 1\% to 33\% of retraining data across tasks. These results suggest that REMEDIS can significantly accelerate the life-cycle of medical imaging AI development thereby presenting an important step forward for medical imaging AI to deliver broad impact.
@misc{azizi_robust_2022,
abstract = {Recent progress in Medical Artificial Intelligence (AI) has delivered systems that can reach clinical expert level performance. However, such systems tend to demonstrate sub-optimal "out-of-distribution" performance when evaluated in clinical settings different from the training environment. A common mitigation strategy is to develop separate systems for each clinical setting using site-specific data [1]. However, this quickly becomes impractical as medical data is time-consuming to acquire and expensive to annotate [2]. Thus, the problem of "data-efficient generalization" presents an ongoing difficulty for Medical AI development. Although progress in representation learning shows promise, their benefits have not been rigorously studied, specifically for out-of-distribution settings. To meet these challenges, we present REMEDIS, a unified representation learning strategy to improve robustness and data-efficiency of medical imaging AI. REMEDIS uses a generic combination of large-scale supervised transfer learning with self-supervised learning and requires little task-specific customization. We study a diverse range of medical imaging tasks and simulate three realistic application scenarios using retrospective data. REMEDIS exhibits significantly improved in-distribution performance with up to 11.5\% relative improvement in diagnostic accuracy over a strong supervised baseline. More importantly, our strategy leads to strong data-efficient generalization of medical imaging AI, matching strong supervised baselines using between 1\% to 33\% of retraining data across tasks. These results suggest that REMEDIS can significantly accelerate the life-cycle of medical imaging AI development thereby presenting an important step forward for medical imaging AI to deliver broad impact.},
added-at = {2023-08-28T10:32:21.000+0200},
author = {Azizi, Shekoofeh and Culp, Laura and Freyberg, Jan and Mustafa, Basil and Baur, Sebastien and Kornblith, Simon and Chen, Ting and MacWilliams, Patricia and Mahdavi, S. Sara and Wulczyn, Ellery and Babenko, Boris and Wilson, Megan and Loh, Aaron and Chen, Po-Hsuan Cameron and Liu, Yuan and Bavishi, Pinal and McKinney, Scott Mayer and Winkens, Jim and Roy, Abhijit Guha and Beaver, Zach and Ryan, Fiona and Krogue, Justin and Etemadi, Mozziyar and Telang, Umesh and Liu, Yun and Peng, Lily and Corrado, Greg S. and Webster, Dale R. and Fleet, David and Hinton, Geoffrey and Houlsby, Neil and Karthikesalingam, Alan and Norouzi, Mohammad and Natarajan, Vivek},
biburl = {https://www.bibsonomy.org/bibtex/27cde2a34e8a7d52e44005563a6049857/serafsoft},
file = {Azizi et al. - 2022 - Robust and Efficient Medical Imaging with Self-Sup.pdf:files/24/Azizi et al. - 2022 - Robust and Efficient Medical Imaging with Self-Sup.pdf:application/pdf},
interhash = {4e8d44d26eae3d3ae46aa3616eba832b},
intrahash = {7cde2a34e8a7d52e44005563a6049857},
keywords = {- Artificial Computer Intelligence, Learning Machine Pattern Recognition, Science Vision and},
language = {en},
month = jul,
note = {arXiv:2205.09723 [cs]},
publisher = {arXiv},
timestamp = {2023-08-28T10:32:33.000+0200},
title = {Robust and {Efficient} {Medical} {Imaging} with {Self}-{Supervision}},
url = {http://arxiv.org/abs/2205.09723},
urldate = {2023-06-13},
year = 2022
}