An evolutionary deep belief network extreme learning-based for breast cancer diagnosis

Cancer is one of the leading causes of morbidity and mortality worldwide with increasing prevalence. Breast cancer is the most common type among women, and its early diagnosis is crucially important. Cancer diagnosis is a classification problem, where its nature requires very high classification accuracy. As artificial neural networks (ANNs) have a high capability in modeling nonlinear relationships in data, they are frequently used as good global approximators in prediction and classification problems. However, in complex problems such as diagnosing breast cancer, shallow ANNs may cause certain problems due to their limited capacity of modeling and representation. Therefore, deep architectures are essential for extracting the complicated structure of cancer data. Under such circumstances, deep belief networks (DBNs) are appropriate choice whose application involves two major challenges: (1) the method of fine-tuning the network weights and biases and (2) the number of hidden layers and neurons. The present study suggests two novel evolutionary methods, namely E-(T)-DBN-BP-ELM and E-(T)-DBN-ELM-BP, that address the first challenge via combining DBN with extreme learning machine (ELM) classifier. In the proposed methods, because of the very large solution space of DBN topologies, the genetic algorithm (GA), which is able to search globally in the solutions space wondrously, has been applied for architecture optimization to tackle the second challenge. The third proposed method in this paper, E-(TW)-DBN, uses GA to solve both challenges, in which DBN topology and weights evolve simultaneously. The proposed models are tested using two breast cancer datasets and compared with the state-of-the-art methods in the literature in terms of classification performance metrics and area under ROC (AUC) curves. According to the results, the proposed methods exhibit very high diagnostic performance in classification of breast cancer.