Abstract
In recent years, an increasing number of women worldwide have been affected by breast cancer. Early detection is crucial, as it is the only way to identify abnormalities at an early stage. However, most deep learning models developed for classifying breast cancer abnormalities tend to be large-scale and computationally intensive, often overlooking the constraints of cost and limited computational resources. This research addresses these challenges by utilizing the CBIS-DDSM dataset and introducing a novel concatenated classification architecture and a two-stage strategy to develop an optimized, lightweight model for breast mass abnormality classification. Through data augmentation and image preprocessing, the proposed model demonstrates a superior performance compared to standalone CNN and DNN models. The two-stage strategy involves first constructing a compact model using knowledge distillation and then refining its structure with a heuristic approach known as Simplified Swarm Optimization (SSO). The experimental results confirm that knowledge distillation significantly enhances the model's performance. Furthermore, by applying SSO's full-variable update mechanism, the final model-SSO-Concatenated NASNetMobile (SSO-CNNM)-achieves outstanding performance metrics. It attains a compression rate of 96.17%, along with accuracy, precision, recall, and AUC scores of 96.47%, 97.4%, 94.94%, and 98.23%, respectively, outperforming other existing methods.