Optimal Deep Learning Architectures for Semantic Segmentation of Blobs from 3D Human MRI Kidney Images

Abstract

In modern medicine, medical imaging has become the backbone for patient s body newlinescreening, disease diagnosis and treatment planning. X-Rays, CT, MRI and Ultrasound newlineare some popular medical imaging modalities in radiology. In our body, the kidneys newlineare the prominent filtering organs, which are responsible for balancing the water and newlinemineral levels. Any abnormalities occurring in kidney functioning will severely affect newlinethe other body organs too. According to recent survey reports, millions of people are newlinesuffering from various kidney diseases. Among them, kidney cancer (renal carcinoma) newlineis chronic and leading to high mortality rate. Physicians are widely using the Computer- newlineized Tomography (CT) and Magnetic Resonance Imaging (MRI) modalities to diagnose newlinekidney diseases. newlineAnalyzing the kidney medical images (i.e CT, MRI etc.) helps in kidney tumor newline(blobs) detection, severity estimation and staging process. Since a decade, deep learn- newlineing models have been widely used in medical image analysis for classification, predic- newlinetion and segmentation activities. Semantic segmentation of the kidney tumors is the newlineprimary activity in kidney cancer diagnosis. Kidney tumors are blob-like structures newlineand their segmentation helps in collecting the statistics about the infected region. For-mer researchers widely used the U-Net, SCNN, FCN and Mask R-CNN as deep learn- newlineing models for medical image segmentation. Although many semantic segmentation newlinemodels exist in deep learning, they are suffering from the kidney tumor segmentation newlineprocess, due to some intrinsic reasons are: i) Morphological Diversity ii) Object Over- newlinelapping iii) Intensity Variance, iv) and Integrated Noise.In order to address the aforementioned issues involved in kidney tumor segmen-tation process, in this research some custom deep learning models are proposed for MRI kidney tumor segmentation process. At first, an optimal and hybrid semantic seg- newlinementation model using the custom U-Net and Mask R-CNN architectures are designed newlinefor kidney and tumor segmentation. Second,