Boosting Breast Cancer Classification from Microscopic Images Using Attention Mechanism

Attention mechanism is one of the foremost later and effective deep learning methods which have shown tremendous performance in detection and classification tasks, including medical imaging. Breast cancer is a primary cause of women's cancer-related morbidity and mortality globally; thus, early identification of this malignancy will help to reduce the number of fatalities. The accurate classification of a mild and deadly cancer in microscopic breast images can give an efficient and relatively low-cost technique for breast cancer early detection. This paper proposes a deep learning model based on an attention mechanism. The proposed attention mechanism derives its input features from pre-trained models and passes its output through a Multilinear perceptron and SoftMax for classification. We trained all models on the ICIAR2018 Grand challenge Breast Cancer dataset and compared them with the pre-trained and ensemble models. For quantitative analysis, validation tests were conducted using the performance metrics for each approach. The suggested approach is proven successful, with classification results improving by +1-6%, potentially reducing human errors in the diagnosis process. Furthermore, the proposed method outperforms the state-of-the-art accuracy, with a +3-8% improvement in performance results.