According to Statistic Canada Website, cancer is one the leading cause of death in this country. The mortality rate is 30% of all causes. Among the different type of cancers, brain cancer is one of the most lethal form due to the fact that brain tumors destroy important brain structures critical to basic body functions. On average, 70% of the patient diagnosed with brain tumor will die. These statistics illustrate the importance of improving the early detection and treatment of brain tumors in order to increase the survival rate of patients. Generally, medical teams study the treatment options for patients based on a set of 3D images to determine treatment options including: surgery, chemotherapy and radiotherapy. For all these options, the localization and the determination of the size and shape of the tumor are vital information that must be estimated from various imaging modalities such as: Positron Emission Tomography (PET), X-ray Computed Tomography (CT), and Magnetic Resonance Imaging (MRI).
During treatments, patients must be imaged several times to track the tumor evolution. Thus, the medical teams must integrate all the information provided by the different image modalities, acquired at different times, and assess via visual inspection the treatment effectiveness. Based on this visual inspection, the team will make decisions about the type of treatment plan. In order to do so, aside from the difficulties associated with the manual alignment of two different scans, there are many artifacts in medical images that make laborious and inaccurate visual inspection. Furthermore, the assessment through visual inspection yield results completely subjective that varies between oncologists. In order to automate this process, many technologies need to be integrated such as: automatic segmentation, non-rigid registration, and normalization of the MRI intensity. There have been many unsuccessful attempts in the past to automate the process of automatically comparing MRI images between treatments. This is due to the fact that in order to quantifying automatically tumor changes over time one must be able to deal with: distortions inherent to MRI imaging that have nothing to do with tumor growth, the lack of anatomic tissue correspondence at different resolutions, and the intensity variations in both volumes due to MRI field fluctuations.
The main goal of my thesis is to produce geometrically accurate 3-D models of brain tumors from MRI volumetric data by comparing them to deformable atlases. The main hypothesis that guides this work is that by comparing two similar deformable atlas models inferred from the raw MRI data one will be able to perform a more precise quantification of brain tumor growths and shrinkages than comparing directly raw MRI data that are frequently at different resolutions requiring an imprecise interpolation process. In order to do so, we will implement a new probabilistic-based atlas deformable model where the atlas shape is estimated from corrected MRI data. In order to do so, we will develop new atlas inference algorithms that avoid 3D-geometric distortions created by image distortion and by MRI field fluctuations. We are also planning to improve data processing speed by implementing these algorithms on low cost Graphic Processing Units (GPU) clusters which are capable of computing speeds in the order of 2 Teraflops. The secondary objective of this thesis is to develop a novel automatic 3D-segmentation algorithm based on deformable probabilistic atlas which will facilitate the estimation of the size and shape of the tumors for diagnosis, conformal radiation therapy, and surgical planning. This project will contribute to the more general objectives of the Brain Tumor Analysis Project spear headed by the Alberta Ingenuity Center for Machine Learning (AICML) and the University of Alberta Cross Cancer Institute (http://webdocs.cs.ualberta.ca/~btap/index.php).