Development of CNN-based model in predicting patients outcomes
Principal Investigator
Name
Shaunagh McDermott
Degrees
M.D.
Institution
Massachusetts General Hospital
Position Title
Radiologist
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-686
Initial CDAS Request Approval
Jun 30, 2020
Title
Development of CNN-based model in predicting patients outcomes
Summary
Standard machine learning algorithms leverage immense data sets to build powerful predictive models. However, the sub-field of few-shot learning [1] considers the problem of predicting many different targets with only a handful of examples of each. This viewpoint is much more suitable to medical imaging, since individual clinical trials may be small, but we can compare data from multiple different trials. In 2017, a major advance [2] triggered a series of computational discoveries that have significantly improved predictive accuracy [3] and enabled the combination of data from disparate sources [4].
We propose to train a prototypical convolutional neural network on a collection of other datasets available from The Cancer Imaging Archive and apply this to the National Lung Screening Trial to predict patient outcomes using finetuning and linear probing [5]. We will then reverse the process by pretraining a model on the NLST data with the hope of improving predictions in the other clinical trials.
References:
Formal:
[1] Wang, Yaqing, et al. "Generalizing from a few examples: A survey on few-shot learning." ACM Computing Surveys (CSUR) (2019).
[2] Snell, Jake, Kevin Swersky, and Richard Zemel. "Prototypical networks for few-shot learning." Advances in neural information processing systems (2017).
[3] Brown, Tom B., et al. "Language models are few-shot learners." arXiv preprint arXiv:2005.14165 (2020).
[4] Wertheimer, Davis, and Bharath Hariharan. "Few-shot learning with localization in realistic settings." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019).
[5] Chen, Mark, et al. "Generative Pretraining from Pixels." International Conference on Machine Learning (2020).
We propose to train a prototypical convolutional neural network on a collection of other datasets available from The Cancer Imaging Archive and apply this to the National Lung Screening Trial to predict patient outcomes using finetuning and linear probing [5]. We will then reverse the process by pretraining a model on the NLST data with the hope of improving predictions in the other clinical trials.
References:
Formal:
[1] Wang, Yaqing, et al. "Generalizing from a few examples: A survey on few-shot learning." ACM Computing Surveys (CSUR) (2019).
[2] Snell, Jake, Kevin Swersky, and Richard Zemel. "Prototypical networks for few-shot learning." Advances in neural information processing systems (2017).
[3] Brown, Tom B., et al. "Language models are few-shot learners." arXiv preprint arXiv:2005.14165 (2020).
[4] Wertheimer, Davis, and Bharath Hariharan. "Few-shot learning with localization in realistic settings." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2019).
[5] Chen, Mark, et al. "Generative Pretraining from Pixels." International Conference on Machine Learning (2020).
Aims
- Applicability of machine learning in predicting patient outcomes
- Few-shot learning
Collaborators
Michael Gee, Massachusetts General Hospital
William Bradley, Massachusetts General Hospital
Karl Knaub, Massachusetts General Hospital