Development of Deep Learning Techniques for Liver and Tumor Segmentation from CT images

Abstract

Liver tumor segmentation is the most prominent and primary step in treating liver newlinecancer and can also help doctors with proper diagnosis and therapy planning. However, it is challenging because of variations in shape, position and depth of tumors and newlineadjacent boundaries with internal organs around the liver. Many deep learning net- newlineworks have shown a good improvement and in particular U-Net based networks laid newlinethe baseline for segmenting the liver and its tumors. Accurate liver tumor segmentation newlinefrom CT images is still a major problem that impacts the diagnosis process. Heteroge- newlineneous densities, shapes and unclear boundaries make liver tumor extraction even more challenging. This thesis provides a useful insight into deep learning based liver tumor segmentation by improving the segmentation accuracy in three different directions. An encoder-decoder-based architecture, which segments liver tumors with a two-step training process is proposed. First, the network segments the liver, and then tumors are extracted from the liver ROIs. We have scaled down the images into different resolutions at each scale and applied normal convolutions along with the dilations and residual connections to capture broad conceptual information without data loss. An overcomplete U-Net to perform liver tumor segmentation jointly using a curriculum learning strategy is proposed. The network has two branches: an overcomplete branch to fine-grade the small structures and an undercomplete branch to fine-grade the high level structures. This combination allows the network to learn all types of tumor artifacts more accurately. We also changed the conventional learning paradigm to curriculum learning where the input images are fed to the network from easy to hard ones to achieve faster convergence. Proposed a deformable encoder-decoder neural network to perform liver segmentation from multi-modality CT images. The deformable convolutions used are newlinethe best suited for shape variant images and yield a good accuracy. newline