Benchmarking Energy and Performance of Deep Learning Models for Breast Cancer Detection
A recent study published on arXiv examined seven deep learning models for breast cancer detection, assessing both their effectiveness and environmental impacts. The models tested included convolutional neural networks, transformers, and hybrid types, utilizing two datasets: Breast Ultrasound and BreakHis at 400X magnification. Researchers calculated the carbon dioxide emissions generated during both the training and inference processes. Results indicated that while EfficientNet and ResNet achieved notable detection rates, they were also associated with high emissions. Alternatively, DeiT-Tiny showed promising performance, whereas DenseNet121 fell short in accuracy. The findings underscore the increasing concerns regarding AI's ecological footprint in medical imaging.
Key facts
- Seven deep learning models were compared for breast cancer detection.
- Datasets used: Breast Ultrasound and BreakHis 400X.
- Architectures include CNNs, transformers, and hybrid models.
- CO2 emissions were assessed during training and inference.
- EfficientNet and ResNet showed strong performance but higher emissions.
- DeiT-Tiny performed competitively on both datasets.
- DenseNet121 achieved lower accuracy.
- Study published on arXiv with ID 2608.09996.
Entities
Institutions
- arXiv