A Cost-Sensitive Convolution Neural Network for Cancer Subgroups Classification

Classification of cancer subtypes is very important task for the diagnosis and prognosis of cancer. In recent years, deep learning methods have gained considerable popularity for this reason; however, it is difficult to determine the structure of the neural network because the function of the deep network depends largely on its structure. In addition, the high number of genes in the gene expression database and the imbalanced data between different classes have a direct effect on the complexity and performance of cancer subgroup classification models. To address the problem of unbalanced data, a convolution neural network (CNN) model using a cost-sensitive strategy is proposed to increase the model's accuracy in identifying minority classes. On the other hand, the fisher ratio technique is used to reduce genes in the preprocessing stage. In techniques the cost-sensitive method, a cost matrix is created based on the distribution of classes, and then this matrix is used in the CNN network cost function step to calculate the amount of error. Two sets of cancer datasets are used to evaluate the proposed method. The results show that selecting the appropriate genes for classification along with the use of cost-sensitive learning can increase the performance of the proposed method compared to the CNN model without selecting the feature and cost-sensitive learning about 11%, 10% and 18% in terms of three criteria of accuracy, recall and precision, respectively.