ProstadeNet: A High Accuracy Fine-Tuned CNN Model for Prostate Cancer Grade Detection from Enhanced Histopathological Images

Abstract

Prostate cancer is the leading malignancy within the bodies of men globally and it is also the fifth most common cause of male cancer mortality. Prostate cancer needs to be diagnosed early in order to reduce mortality and give better treatment to concerned individuals. Although traditional methods of diagnosing prostate cancer are time-consuming, computer-aided diagnosis systems provide a quicker diagnosis without any loss of accuracy. There is a very alarming need for fast, accurate, and economical diagnostic processes of this epidemic. The main goal of this research is to recommend an effective framework for grading classification of prostate cancer from histopathological images with six classes with fewer errors of classification. Various image augmentation preprocessing techniques have been utilized for improving the quality of images, and four different augmentation techniques to enhance the dataset size and reduce model overfitting problems. Finally, a novel ProstadeNet architecture which is a tuned CNN tuned in structure and hyperparameters by ablation study will be presented. The ProstadeNet model achieved an accuracy of 99.83%. To estimate the capacity of ProstadeNet, relative experiments with five typical transfer learning models will be conducted to find out its relative performance. K-fold validation will be used in measuring results. Various performance measurements have been used to show the robustness of the models.