Abstract

Medical image segmentation is a process of isolating or identifying an object of interest in medical images. It plays a pivotal role in clinical diagnostics, monitoring and treating diseases. The convolutional neural network, U-Net, was specifically developed for segmentation of medical images in view of its ability to accurately segment with limited training data. Existing U-Net based segmentation networks suffer from high computational complexity in order to provide a reasonable performance. This thesis presents five U-Net based schemes that significantly reduce the computational complexity without compromising the performance. In the first part, we develop a number of segmentation schemes referred to as MAGNet, MedSegNet, SSNet, and FFNet that utilize attention mechanism-enhanced multiscale feature fusion. The first two networks are developed for segmenting CT, colonoscopy, and non-mydriatic 3CCD images. SSNet is a semi-supervised network that effectively makes use of both labeled and unlabeled data for segmenting brain anatomical structures of tissues in MR images. FFNet is proposed for segmenting benign and malignant tumors from ultrasound images. In the second part, we present a lightweight attention-guided network with feature recalibration, referred to as LASegNet. The main idea used in designing LASegNet is on one hand to reduce the number of parameters by cutting on the number of filters used and on the other hand, restore the performance by combining the features of the encoder and decoder units through a judicious use of attention guided module. Extensive experiments are performed to demonstrate the effectiveness of each of the schemes proposed. Specifically, it is shown that LASegNet is robust across images from different modalities.