Enhancing and Nonenhancing 3D Brain Tumor Segmentation with Modified Swish Activation and Double U-net Architecture

A Brain tumor is a growth of irregular cells in the brain. (Artificial Intelligence) AI can be used for fully automatic extraction of the tumor region from (Magnetic Resonance Imaging) MRI scans. Developing such a model will decrease the dependency on the radiologist's experience and provides a faster way to visualize the tumor without the time-consuming process of manually segmenting the tumor regions from 3 Dimensional MRI scans. Diagnosing the extent of the tumor region in the early stages can be lifesaving. Automated segmentation models can also be integrated with Medical Imaging devices. In this paper, we use a Deep Convolutional Neural Network inspired by the U-Net architecture for the segmentation of a brain tumor. We propose a Double U-Net architecture with a new, custom activation function, modified Swish function. Our proposed model achieves better performance metrics than the U-Net model with the well-known activation function, ReLU, while having to train on half as many parameters. Our proposed network is trained and evaluated on the Decathlon 10 Challenge dataset, which contains multimodal 3D MRI scans, labeled as enhancing tumor and non-enhancing tumor by experts. In this paper, we also propose a new activation function, i.e., Modified Swish Function(MSW).