Related Experiment Videos
Effect of Deep Learning Training Policy on Greenhouse Gas Emissions and Carbon Efficiency for Chest Radiograph
Nicholas Dietrich1, David McShannon2, Merel Huisman3
1Department of Medical Imaging, Temerty Faculty of Medicine, University of Toronto, ON, Canada.
Summary
Deep learning training policies significantly impact radiology's environmental footprint. Early stopping in AI model training reduces carbon dioxide equivalent (CO2eq) emissions by up to 38% while maintaining performance.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence
- Environmental Sustainability
Background:
- Environmental sustainability is a growing concern in healthcare, particularly within radiology.
- The carbon footprint of deep learning (DL) models in medical imaging is not well understood.
- Quantifying the impact of DL training policies on greenhouse gas emissions is crucial for sustainable practices.
Purpose of the Study:
- To quantify the effect of different DL training policies on carbon dioxide equivalent (CO2eq) emissions.
- To evaluate the impact of these policies on model performance for chest radiograph classification.
- To compare the environmental cost and efficiency of various training strategies.
Main Methods:
- Trained three ImageNet-pretrained convolutional neural networks (ResNet-50, DenseNet-121, EfficientNet-B0) on a large dataset of chest radiographs.
- Evaluated three training policies: retrospective optimal checkpoint selection, prospective early stopping, and fixed 20-epoch training.
- Assessed per-epoch CO2eq emissions, energy consumption, macro-averaged area under the curve (AUC), and carbon efficiency.
Main Results:
- The optimal checkpoint for model deployment often occurred early in training, with significant emissions generated post-optimum (up to 84%).
- Prospective early stopping achieved performance comparable to retrospective optimal checkpoint selection.
- Early stopping reduced total emissions by up to 38% and improved carbon efficiency by up to 76% compared to fixed-epoch training.
Conclusions:
- A substantial portion of DL training emissions can occur after the model has reached its optimal performance.
- Prospective early stopping is an effective strategy to reduce the environmental impact of DL in radiology.
- Implementing early stopping in DL training can lead to significant reductions in greenhouse gas emissions and improved carbon efficiency without compromising diagnostic performance.