A Robust Ensemble Deep Learning Approach for Breast Cancer Diagnosis

Globally, the number one cause of death for women is breast cancer, which is a serious public health concern. Deep learning has shown promising results in detecting and diagnosing breast cancer from medical images. While deep learning models have great potential, identifying the most effective deep learning architecture presents a significant challenge. This paper proposes a two-stage ensemble classification approach for breast cancer detection. The first stage involves training multiple baseline models based on CNN architectures, including VGG16-SVM, ResNet50, DenseNet169, MobileNetV2, and InceptionV3. In the second stage, three different ensemble fusion methods, including voting, weighted voting, and meta-learning, were applied to combine the predictions of the baseline models. The second stage was introduced to overcome the challenge of selecting the best deep learning model by fusing a set of ensemble techniques. In the second stage, four different ensemble fusion methods are used, including hard voting, soft voting, weighted voting, and meta-learning. The findings of the experimentation on IHC images demonstrate that the fusion techniques have led to a boost in performance in comparison to the baseline CNN-based models. Furthermore, implementing an ensemble approach utilizing meta-learning has exhibited notable potential in augmenting the overall performance, Due to the imbalance in the dataset, the performance is evaluated in terms of F1-score. Its performance achieved F1-score of 89.2% surpassing the best baseline deep learning models by 22.2%.