A new approach for data augmentation in a deep neural network to implement a monitoring system for detecting prostate cancer in MRI images1

Abstract

Prostate cancer is one of the most common cancers in men, which takes many victims every year due to its latent symptoms. Thus, early diagnosis of the extent of the lesion can help the physician and the patient in the treatment process. Nowadays, detection and labeling of objects in medical images has become especially important. In this article, the prostate gland is first detected in T2 W MRI images by the Faster R-CNN network based on the AlexNet architecture and separated from the rest of the image. Using the Faster R-CNN network in the separation phase, the accuracy will increase as this network is a model of CNN-based target detection networks and is functionally coordinated with the subsequent CNN network. Meanwhile, the problem of insufficient data with the data augmentation method was corrected in the preprocessing stage, for which different filters were used. Use of different filters to increase the data instead of the usual augmentation methods would eliminate the preprocessing stage. Also, with the presence of raw images in the next steps, it was proven that there was no need for a preprocessing step and the main images could also be the input data. By eliminating the preprocessing step, the response speed increased. Then, in order to classify benign and malignant cancer images, two deep learning architectures were used under the supervision of ResNet18 and GoogleNet. Then, by calculating the Confusion Matrix parameters and drawing the ROC diagram, the capability of this process was measured. By obtaining Accuracy = 95.7%, DSC = 96.77% and AUC = 99.17%, The results revealed that this method could outperform other well-known methods in this field (DSC = 95%) and (AUC = 91%).