ICMLA 2026 study: DeiT-Tiny best balances accuracy and CO2 for breast cancer AI
Comparing 7 deep learning models, DeiT-Tiny cuts emissions while keeping accuracy on ultrasound data.
A new study from researchers Samar Garrab and Ghada Achour, accepted at ICMLA 2026, benchmarks seven deep learning architectures for breast cancer detection, comparing both diagnostic performance and environmental cost. The models span CNNs, transformers, and hybrids, and were evaluated on two medical imaging datasets: Breast Ultrasound and BreakHis 400X. The authors measured CO2 emissions during both training and inference, moving beyond accuracy-only comparisons to address the growing carbon footprint of complex AI models.
Key findings reveal a clear tradeoff. EfficientNet and ResNet delivered strong accuracy but carried the highest CO2 emissions, while DenseNet121 lagged in accuracy. Among transformers, DeiT-Tiny proved surprisingly competitive, achieving the most favorable balance between accuracy and energy consumption on the Breast Ultrasound dataset. On BreakHis, however, ViT and Swin transformer models produced the best results. The authors conclude that no single model category consistently wins across datasets, emphasizing that model selection for medical applications must jointly factor in performance, emissions, and dataset-specific characteristics.
- EfficientNet and ResNet achieve top accuracy but generate the highest CO2 emissions during training and inference.
- DeiT-Tiny delivers the best accuracy-to-energy tradeoff on the Breast Ultrasound dataset, while ViT and Swin lead on BreakHis 400X.
- Study across 7 architectures and 2 datasets found no single model family dominates, urging emissions-aware model selection in medical AI.
Why It Matters
Medical AI teams can now pick models that cut carbon footprint without sacrificing diagnostic accuracy, per dataset.