- Keywords: medical imaging, deep learning, carbon footprint
- TL;DR: Carbon footprint driven deep learning model selection for medical imaging.
- Abstract: Selecting task appropriate deep learning models is a resource intensive process; more so when working with large quantities of high dimensional data that are encountered in medical imaging. Model selection procedures that are primarily aimed at improving performance measures such as accuracy could become biased towards resource intensive models. In this work, we propose to inform and drive the model selection procedure using the carbon footprint of training deep learning models as a complementary measure along with other standard performance metrics. We experimentally demonstrate that increasing carbon footprint of large models might not necessarily translate into proportional performance gains, and suggest useful trade-offs to obtain resource efficient models.
- Paper Type: validation/application paper
- Primary Subject Area: Application: Other
- Secondary Subject Area: Validation Study
- Paper Status: original work, not submitted yet
- Source Code Url: https://github.com/lfwa/carbontracker/
- Data Set Url: https://wiki.cancerimagingarchive.net/display/Public/LIDC-IDRI
- Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
- Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.