Attention-Based End-to-End Hybrid Ensemble Model for Breast Cancer Multi-Classification

Abstract

The daily rise in female instances of breast cancer (BC) is largely due to misinformation and late-stage detection. Effective treatment for BC can only be administered by correctly diagnosing cancer in its very early stages of development. BC classification has been discovered to be accelerated and automated using deep learning models and medical image analysis techniques. However, these techniques, which are crucial for numerous other applications outside broad visual identification, usually use generic traits. Since the primary objective of deep learning models is to characterize complex boundaries of hundreds of classes in deep space, embracing higher-order qualities is essential for improving non-linear modeling abilities. This study employs the publicly available BreakHis to provide an end-to-end hybrid ensemble model for BC multi-classification utilizing an attention-based global second-order pooling network. Ensembling is accomplished by adding an attention-based second-order pooling network in the form of a convolutional layer to the separate models to increase their non-linear modeling skills before concatenating their output features. Finally, the output features are relied on a classification layer for the final forecast. The proposed model produced enhanced results for binary and multiclass (four classes and eight classes) classification with 97.6% (40x), 95.5% (100x), 96.6 (200x) and 95.9% (400x) accuracy for the eight classes experiment. The experimental results show that, when compared to state-of-the-art models, the proposed approach obtains the best BC multi-classification accuracy.