Performance Study of Optimizers for Segmentation of Brain Tumors using Atrous Convolution in U-Net

Abstract

The purpose of segmentation of the tumourous region in the brain is to identify the tumorous brain tissues and their location. Manually segmenting the brain tumor is not only time-consuming and error-prone but also increases the rate of mortality. In the time of need for quick and precise segmentation, many convolution neural networks have shown exceptionally good execution. In medical image segmentation, the most favored network that has been used is U-Net. In this work, reinstatement of the 2D convolution layer with a 2D dilated convolution layer or atrous convolution layer with a dilation rate as 2 has been implemented. The performance has been analyzed during the training and testing of the model. Different optimizers like Stochastic Gradient Descent (SGD), SGD along with momentum, Adam, and Nadam (Nesterov Adam) have been used in the U-Net model. MICCAI BraTS 2020 dataset has been used in this work. From the simulated results, it has been observed that the Nadam optimizer along with the atrous convolution layer provides higher accuracy