Computational Efficiency and Accuracy of Deep Learning Models for Automated Breast Cancer Detection in Ultrasound Imaging
Luaay Alswilem, Nurettin Pacal
Abstract
This study explores the trade-off between diagnostic performance and computational efficiency in deep learning models for the classification of breast cancer in ultrasound images. To this end, we evaluate three contemporary CNN architectures EfficientNetB7, EfficientNetV2-Small, and RexNet-200 in a multiple comparative study with standardized performance and complexity metrics. Our evaluations provide evidence that all three models achieved an identical high accuracy of 95.00%, but there were sizeable differences in the computational resources required to achieve that accuracy. RexNet-200 demonstrated tremendous computational efficiency, achieving identical performance with the least amount of resources (13.81M parameters; 3.05 GFLOPs) required compared to EfficientNetB7 which is much more computationally intensive. An examination of the confusion matrix for the models enhances the models clinical validity, as there are no malignant lesions misclassified as normal. Ultimately, our study clearly demonstrates that diagnostic accuracy is not a good metric for practical clinical deployment. RexNet-200, by representing high performance, with minimal resource utilization, is the most pragmatic and clinically applicable model, creating the opportunity to develop scalable and accessible CAD systems in resource-limited settings.