Deep Neural Networks for segmentation of CT medical images

In the last few years, the Artificial Intelligence (AI) has made big steps forward in many fields of telemedicine, especially in the process of biomedical images acquired by several technologies and techniques like Ultrasound, Computed Tomography (TC), Magnetic Risonance Imaging (MRI), etc.. In particular, the segmentation of biomedical images is very important to identify pathologies and lesions of organs allowing the doctor to intervene before any complications of the organ disease. This thesis project is developed at SynDiag srl, a company that develops artificial intelligence algorithms to support the interpretation of medical images, in particular ultrasound images of ovarian cysts. The goal is the automatic segmentation, through the implementation of Deep Neural Networks (DNN) like the U-net, of CT images concerning the liver, using the software library of Tensorflow 2.0. The U-net architecture is applied on a database composed of 31 clinical cases, for a total of 2898 images in the train dataset and 1242 images in the test dataset. Due to the small amount of the training images, the U-net is implemented also on a larger training dataset by applying the data augmentation. To do this, several augmentation techniques have been tested in order to identify those that improve the average score of the network.