U^3-NET: NESTED CONVOLUTIONAL NEURAL NETWORK FOR BIOMEDICAL IMAGE SEGMENTATION

Abstract

Understanding biological processes and analyzing different diseases requires accurate segmentation of the biomedical images that researchers have available. In diseases like cancer, it can help in developing drugs rapidly and applying proper treatments. However, labeling all the objects and drawing contours around them in an MRI image or CT scan requires knowledge and experience. Fortunately for all of us, artificial intelligence algorithms in computer vision have advanced a lot and many tasks in image processing can be solved using Convolutional Neural Networks. In this thesis, we designed a deep and powerful architecture, U^3-Net, for

biomedical image segmentation. The architecture of our U^3-Net network is a three- level nested U-Net-like shape. Our proposed architecture can be implemented from

scratch without using image classification backbones. Nuclei segmentation and finding cell confluence were the successful tasks that we solved using U^3-Net. Finally, we measured the evaluation metrics of our proposed model and compared it to other state-of-art architectures widely used in medical image segmentation. In most of the datasets that we evaluated, U^3-Net achieved the highest metrics.