Deep Learning-Based Semantic Segmentation Models for Prostate Gland Segmentation

Abstract

Prostate cancer (PCa) is one of the prevalent forms of cancer disease found in males due to the unusual development of cells. Early diagnosis of this PCa can be useful in terms of treatment and medication. Segmentation and classification of the PCa through manual observation are one of the diagnosis methods, but it is highly challenging due to complex boundaries and features. Machine learning-based semantic segmentation architecture models consume more energy and processing time and will lead to reduced scalability and reliability. In order to tackle these limitations, deep learning-based semantic segmentation architecture can be used as it has more advantages in discriminating the features of the lesions efficiently and accurately. The main aim of this work is to segment the PCa lesions accurately and efficiently. Hence, deep learning-based semantic segmentation model-based architectures such as U-Net, Linknet, and PSPNet are proposed in this research. These models are equipped with a backbone as Inception-ResNet-v2 CNN architecture for prostate cancer gland segmentation. Nearest neighbour interpolation and normalization methods are employed as the preprocessing technique for enhancing the PCa MRI images. The normalized image was taken for processing various settings of U-Net, LinkNet and PSP-Net architectures for performing segmentation and for optimizing these models, Adam, Adamax and Nadam optimizers are used. The experiment was performed using NCI-ISBI 2013 dataset. Performance analysis of the proposed models is evaluated using Intersection of Union (IoU) scores, where the LinkNet optimized with Adamax obtained a best IoU score of 0.763337802.