Vascular segmentation of brain MR angiography images using convolutional neural networks

Segmentation of vascular structures is important for constructing 3D models of the vessel anatomy and for detecting vascular anomalies in patients. The constructed models of normal anatomy and anomalies are used in surgical planning, navigation and monitoring of patients in related diseases. The goal is to obtain improved diagnosis, intervention and treatment follow-up in vascular diseases. In this thesis, we developed a novel vessel segmentation method to obtain 3D models of the cerebral arteries using Magnetic Resonance Angiography data. Saccular aneurysms and alterovenous malformation which are some of these aforementioned anomalies, have high mortality and morbitity rates. Time-of-Flight MRA, which is the MRA procedure used to detect these anomalies provides the vessel structure to appear more distinct and eases the process. Previously, convolutional neural networks (CNN) obtained great success in competitions with broad participation like ILSVRC and atracted the researchers studying in areas like semantic segmentation, classification and image retrieval. This trend recently also appeared in medical image classification and segmentation. Many successful studies are reported in domains such as brain lesion segmentation, brain tumor segmentation and lung lesion segmentation. However, currently no solution exists for vascular segmentation of the brain. Three different CNN architectures are adopted for vessel segmentation in this thesis: MODEL1, MODEL2, and a three pathway 2D architecture originally used for tissue segmentation. From these architectures, first two are our contributions and the latter is the architecture which was used in recent studies for a different purpose. We also evaluated different combinations of these architectures' results. The architectures presented in this thesis are tested in two different datasets. First data set was gathered from MIDAS MRA data set. As pre-processing, using a brain extraction method, structures like skull and nasal bone are eliminated. After that using a window/level transformation, a homogeneous distribution of color intensities between 0-255 is ensured. We also create and use a synthetic vessel dataset. For this data set, statistics of orientation and radius of vessel structures in MIDAS data set are obtained and new images are created randomly using these statistics. Consequently, a dataset with more curved structures and larger variations is obtained. Using both the synthetic and the real MRA datasets, the new CNN architectures are trained and tested. The obtained results of %68 dice ratio in the test set, demonstrate the capability of CNNs in the vessel segmentation problem, which is more challenging than other applications of segmentation, due to highly vascular anatomy, complex geometry and topology of tubular and bifurcating structures in vessels, and pathologies such as aneurysms. Furthermore, in this thesis, using segmentation results obtained from different architectures, fusion segmentations are obtained. For this purpose, as the first method, intersections of positive predictions is acquired and using the results from all the three architectures, a Dice ratio of %52 is obtained. As another method, weighted majority voting is used in which, different classifiers contributes to the final prediction according to their performances but no improvement is obtained.