Effect of Deep Learning Training Policy on Greenhouse Gas Emissions and Carbon Efficiency for Chest Radiograph Classification
Nicholas Dietrich, David McShannon, Merel Huisman, Florence X. Doo, Kate HannemanPurpose:
Environmental sustainability is an emerging priority in radiology, yet the impact of deep learning training policies on greenhouse gas emissions remains poorly characterized. This study quantified the effect of training policy on carbon dioxide equivalent (CO 2 eq) emissions and model performance for chest radiograph classification.
Methods:
Anteroposterior chest radiographs (128 907 training, 24 570 validation, 8282 test) were used to train 3 ImageNet-pretrained convolutional neural networks (ResNet-50, DenseNet-121, EfficientNet-B0) for 20 epochs. Three policies were evaluated: (1) retrospective optimal checkpoint selection at the validation loss minimum; (2) prospective early stopping (patience 10 epochs); and (3) fixed 20-epoch training without checkpoint selection. Per-epoch CO 2 eq emissions, energy, macro-averaged area under the curve (AUC), and carbon efficiency were evaluated.
Results:
Validation loss reached its minimum at median epoch 2 for ResNet-50 and DenseNet-121 and epoch 4 for EfficientNet-B0. At the retrospective optimum, macro-AUCs ranged from 0.793 to 0.800 and generated 6.2 to 7.9 g CO 2 eq (37-46 Wh) at the deployed checkpoint. However, producing this model required the full run, generating 30.8 to 49.4 g CO 2 eq (181-291 Wh) with 78% to 84% of training emissions accruing after the deployed checkpoint. Prospective early stopping had macro-AUCs equivalent to the retrospective optimum (0.793-0.800), with 31% to 38% lower total emissions (21.1-30.9 vs 30.8-49.4 g CO 2 eq) and 57% to 76% higher carbon efficiency (25.9-37.6 vs 14.7-24.0 AUC/kg CO 2 eq) compared to fixed-epoch training.
Conclusions:
Up to 84% of total training emissions accrued after the optimal checkpoint, with relative savings dependent on the comparator. Prospective early stopping preserved performance, reduced emissions by up to 38%, and improved carbon efficiency by up to 76% versus fixed 20-epoch training.