Optimal Parameter Selection-Based Deep Semi-Supervised Generative Learning and CNN for Ovarian Cancer Classification

A segmentation and categorization of ovarian cancer varieties from Computed Tomography (CT) scans is greatly necessary in current medicinal diagnosis system to lessen the mortality rate. To perform this task, a Deep Semi-Supervised Generative Learning with Enhanced U-Net and fused Deep Convolutional Neural Network (DSSGL-EUNet-DCNN) was developed to augment the training ovarian CT scans, partition the Region-Of-Interests (ROIs), and classify the varieties of ovarian cancer. But, its efficiency depends on the selection of hyperparameters for learning the deep learner. Hence in this article, a DSSGL-EUNet with Multi-Scale DCNN (DSSGL-EUNet-MSDCNN) model is proposed which contains different kernel sizes, learning rate and batch size for multiple DCNN to classify the types of ovarian cancers. First, the training CT scans are augmented by the DSSGL and the ROIs from each CT scan are segmented by the EUNet models. Then, the segmented ROIs are fed to the fused DCNN structure in which every DCNN captures the features from each segment at a scale-level. Also, the hyperparameters of DCNNs are chosen by the lion optimization algorithm for feature extraction and classification. Based on this process, the training errors and time cost are reduced with high classification accuracy. At last, the experimental results exhibit that the DSSGL-EUNet-MSDCNN realizes a higher accuracy than the classical models for segmentation and classification of ovarian cancers.