Generative Adversarial Networks for 3D Medical Image Translation
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2021-08-23
Department
Major/Subject
Machine Learning, Data Science and Artificial Intelligence
Mcode
SCI3044
Degree programme
Master’s Programme in Computer, Communication and Information Sciences
Language
en
Pages
65
Series
Abstract
Medical image-to-image translation has the potential to reduce the imaging workload at clinics, by removing the need to capture some sequences. In addition, it has the potential to reduce the annotation burden, class imbalances and lack of data for developing machine learning methods, such as classification and segmentation models. A family of generative models, called Generative Adversarial Networks (GANs), have been used successfully to translate images from one domain to another, such as MR to CT. The aim of this thesis is to investigate the performance of different GAN models on medical image-to-image translation tasks, including MR to CT, CT to MR, and MR sequence to another MR sequence. The work is performed using a dataset of paired (voxel-wise aligned) Dixon MR sequences, including Dixon-Fat (DF), Dixon-In-Phase (DIP) and Dixon-Water (DW), and Magnetic Resonance for Calculating ATtenuation (MRCAT). MRCAT is a synthetic CT scan produced commercially by Philips. In addition, we use a dataset of CT scans that are not paired with the other data. We train GANs for paired and unpaired image-to-image translation, including pix2pix and pix2pixHD (paired) and CycleGAN (unpaired) as baseline models for different image-to-image translation tasks with the available data. The first main goal of this thesis is to develop image-to-image translation models for tasks MR-CT and CT-MR with unpaired MR and CT data. For this goal we developed pix2pix variants, called pix2pixM-C and pix2pixC-M, that utilize unpaired CT and MR data, and MRCAT pairs for the MRs. The second goal is to utilize a multi-domain model that learns mappings between many different domains, making image-to-image translation more scalable. For this we utilize a StarGAN v2 variant, StarGAN v2pair, with an additional L1 loss between the generated synthetic images and the ground truth images. We compare the performance of the models with the baselines. pix2pixM-C and pix2pixC-M outperformed the baselines on MR-CT and CT-MR tasks, respectively, in terms of Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) scores, while pix2pixHD outperformed the other models on MR1 to MR2 translation tasks, where MR1 and MR2 are two different MR sequences.Description
Supervisor
Kannala, JuhoThesis advisor
Akram, SaadKeywords
image-to-image translation, generative adversarial networks, deep learning, computer vision, machine learning, medical image analysis