Deep learning based liver segmentation from computed tomography images

Nowadays, when there are high mortality rates due to liver-related diseases, it is of great importance to make the correct diagnosis in a short time with medical imaging methods. Although this task is performed by radiologists and oncologists, automating the diagnosis with computer-aided diagnostic methods will provide serious benefits in terms of shortening the diagnosis time, preventing differences in interpretation between specialists, and rapid transition to the treatment phase. For this reason, studies on computed tomography (CT) images, which is one of the medical imaging methods, are gaining momentum and especially separating the liver from other organs (segmenting) from abdominal CT images prepares the ground for lesion detection and classification in the liver in the second stage. Although there are many methods for segmenting CT images, convolutional neural network (CNN) structures, which are widely used today, leave other methods behind with high success rates. Training CNN structures to perform the given task requires very large data sets and high-performance hardware and software. For this reason, the method of training pre-trained CNN structures with a smaller data set for the task to be used, called transfer learning, will reduce the length of training time with large data sets and the financial burden of the source software and hardware where learning will take place. In this thesis study, the VGG16, VGG19, RESNET18, RESNET50 models, which are CNN structures pre-trained with the large dataset called ImageNet, using transfer learning method, will be retrained with the open dataset containing LiTS2017 abdominal CT images and their performance in segmenting the liver in image sections will be compared.