Improvement U-Net for MRI Brain Tumour Segmentation by Searching for Suitable Activation Function

Abstract

This study proposes the usage of enhanced activation function to improve the performance of deep learning models used in MRI brain tumour segmentation. Activation function has a significant role in deep network stability, learning rate and accuracy of the resultant solution. The main advantage of this new activation function is the ability to provide more accurate results solving the problem of vanishing gradient in comparison to the common activation functions “ReLU”. Vanishing gradient affects the training rate and therefore, the weight updates and the overall network accuracy. This work aims to study the feasibility of increasing the accuracy of deep learning brain tumour segmentation using an enhanced activation function. In this study, U-Net deep learning model was chosen. A modified U-Net architecture was built for the segmentation task. Several enhanced activation functions were used besides the standard ReLU. The benchmark database used for the evaluation was BRAST 2015 dataset. The results showed that the HardELiSH activation function outperformed the standard ReLU activation function. This proves that the deep learning model performance in brain tumour segmentation can be enhanced with the choice of the activation function.