Low-Complexity Low-Memory VGG Models for Accurate Diagnosis of Breast Cancer
Md. Bipul Hossain, Mohamed Shaban · 2024
Breast cancer is considered the most prevalent cancer among women and may spread to other organs if it is not detected and treated early. In addition, mammography is the primary imaging modality and is widely used for screening and diagnosing breast cancer. Recent studies have applied deep learning for mammogram-based breast cancer detection. However, these methods have high computational and memory requirements. In this paper, we propose low-complexity light-weight VGG-16 and VGG-19 models that are able to detect breast cancer at an accuracy of 98% and a sensitivity of 95%. While the performance of the proposed models is comparable to that of the-state-of-the-art architectures, the proposed models were extensively compressed using a novel combination of filter pruning, node pruning and post-training quantization offering up to 98.4% reduction in the number of learning parameters, up to 97.6% drop in the number of FLOPs, and almost 99.6% decrease in the allocated memory. This will potentially facilitate the efficient implementation of the models on edge Internet of Things (IoT) devices deployed in future IoT healthcare platforms.