Enhancing Breast Cancer Classification using a Modified GoogLeNet Architecture with Attention Mechanism

Abstract

Breast cancer incidence has been soaring sharply, and this is causing grave concern worldwide due to its high mortality rates. It should be correctly diagnosed in the early stages in order to achieve better patient outcomes. Over the last decade, there has been a great demand for diagnosis systems based on AI that could be used in breast cancer detection and classification. These computerized devices utilize deep learning algorithms to analyze medical scans thereby allowing for subtle abnormality recognition and distinguishing malignant from benign tumors. Computer-aided diagnosis named CAD systems can assist radiologists and pathologists to be more precise with their diagnoses while at the same time increasing productivity. Furthermore, recent advances in CNN architectures coupled with attention mechanisms have further improved CAD systems for breast cancer diagnosis. Attention-based CNN models focus on crucial regions hence enhancing classification accuracy and reliability. In this study, we introduce a new approach that improves the classification of breast cancer using GoogLeNet architecture modified by an attention mechanism that is based on image regions. The modified GoogLeNet has a spatial transformer network (STN), which allows it to focus on significant areas in breast histopathology images using selective attention. Through the attention mechanism, the model becomes better at learning discriminatory features that indicate different subtypes of breast cancer. In order to evaluate the effectiveness of this method, we implemented experiments using the BreaKHis dataset for classifying breast carcinomas. This dataset has been intentionally collected under various magnifications so as to facilitate binary and multiple classification tasks. These outcomes clearly show that modified GoogLeNet with attention outperforms the original GoogLeNet architecture in terms of accuracy. For binary classification, the proposed model demonstrated an accuracy rate of 98.08%, whereas GoogLeNet’s rate was 94.99%. For multi-class classification, at 100x magnification, this model achieved an accuracy of 94.63% while the accuracy of the original GoogLeNet was 85.06%. It is evident from these findings that the efficiency of breast cancer diagnosis significantly improved under this proposed approach. The findings of the study show that incorporating the modified GoogLeNet framework with an attention mechanism in CAD systems for breast cancer detection can improve their performance. Combining deep learning and attention models together can lead to more accurate treatment decisions and better patient results. More efforts are needed to further develop CAD systems in this area to assist ongoing endeavors towards upgrading them and ultimately contribute to saving many lives in the fight against breast cancer.

Keywords

• Breast cancer, computer-aided diagnosis, deep learning, attention mechanism, spatial transformer network