acegreenhouse gas emissions and carbon efficiency of chest radiograph classification training under different stopping policies
Source article: Effect of Deep Learning Training Policy on Greenhouse Gas Emissions and Carbon Efficiency for Chest Radiograph Classification
Abstract: Purpose 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,…
Positive state: both sides are scored from claims and sources, not community votes.
Office of Government Policy Coordination South Korea by Minseong Kim. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0
In a study published August 12, 2026, researchers quantified CO2eq emissions for training ResNet-50, DenseNet-121 and EfficientNet-B0 on 128,907 chest radiographs for 20 epochs. They found validation loss minima at median epochs 2 to 4, meaning most emissions occurred after the best checkpoint, and compared retrospective selection, prospective early stopping, and fixed-epoch training on AUC and energy use.
The findings matter because radiology AI development carries a measurable carbon cost that can be reduced without sacrificing diagnostic performance. Prospective early stopping preserved macro-AUC around 0.793-0.800 while lowering total emissions by 31% to 38% and improving carbon efficiency by 57% to 76% versus fixed training, though savings depend on comparator and the analysis is limited to three architectures and a single chest X-ray task.
- Study used 128,907 training, 24,570 validation, and 8,282 test anteroposterior chest radiographs to train ResNet-50, DenseNet-121, and EfficientNet-B0 for 20 epochs.
- Validation loss minimum occurred at median epoch 2 for ResNet-50 and DenseNet-121 and epoch 4 for EfficientNet-B0.
- At retrospective optimum, models generated 6.2 to 7.9 g CO2 eq (37-46 Wh) at deployed checkpoint but required full run of 30.8 to 49.4 g CO2 eq (181-291 Wh).
Prospective early stopping with patience 10 preserved chest radiograph classification performance while lowering total training emissions by up to 38% and raising carbon efficiency by up to 76% compared to fixed 20-epoch training.
Fixed 20-epoch training without checkpoint selection wasted most compute, with 78% to 84% of total emissions occurring after the optimal checkpoint had already been reached.
The rundown
Researchers trained three ImageNet-pretrained CNNs on anteroposterior chest radiographs and compared three policies: retrospective optimal checkpoint selection at validation loss minimum, prospective early stopping with patience 10, and fixed 20-epoch training. They tracked per-epoch CO2eq, energy in Wh, macro-averaged AUC, and carbon efficiency.
Retrospective optimum macro-AUCs ranged from 0.793 to 0.800 with only 6.2 to 7.9 g CO2eq at the deployed checkpoint, yet the full run needed to find it emitted 30.8 to 49.4 g. Prospective early stopping matched retrospective performance while cutting total emissions to 21.1-30.9 g and increasing efficiency to 25.9-37.6 AUC/kg CO2eq.
Sources
- Peer-reviewedCanadian Association of Radiologists Journal2026-08-12
How should this claim be treated?
The debate