Investigation of different deep learning techniques in pancreas cancer tissues segmentation

Automatic pancreas and pancreatic tumor segmentation is used to prevent late diagnosis of pancreatic cancer and to assist medical doctors in diagnosis, treatment and surgery. Because of the variable size, shape and location of the pancreas and pancreatic tumor, studies on pancreas and pancreatic tumor segmentation can achieve a certain percentage of success. The aim of our thesis is to provide automatic pancreas and pancreatic tumor segmentation with higher accuracy in CT imaging. In this context, deep learning-based approaches are recommended. Our thesis consists of two different parts. The first part proposes a two-stage method for performing pancreatic segmentation; (i) Determination of Pancreas Region of Interest and (ii) Pancreas Segmentation. In the first stage, it roughly determines the position of the pancreas. This step produces 2D sub-CT slices of candidate regions of the segmented and masked pancreas. The second stage (Pancreas Segmentation) takes the 2D sub-CT slices produced in the previous stage as input and the segmented pancreas region is produced as output. In the second part of the thesis, the two-phase method proposed in the first part is redesigned to segment pancreas and pancreatic tumor tissues and improvements are made in the performance of each phase. The proposed approaches for each part are compared with the studies in the literature and it is proved that they provide better results subjectively and objectively.