Point and pattern detection in 4D PC-MRI blood flow

Cardiovascular diseases (CVDs) are the number one cause of death worldwide[1]. It is important to investigate these diseases in order to find better diagnostic techniques. Blood-flow data provides important information for the analysis of cardiovascular diseases, since the bloodstream influences vessel walls, and vice versa. Anomalies in the blood flow can therefore be a result or cause of CVDs. Blood-flow measurements can reveal patterns and anomalies, and can be used to gain a better understanding of the cardiovascular physiology and pathology. Visual inspection of blood-flow data potentially provides newfound insight in the flow characteristics. However, visual analysis is challenging because of the abundance and complexity of information. New visualization methods showing clearly the anomalous flow areas are needed. This often involves simplification of the velocity fields by means of feature extraction. The goal of this research is to compare and cross-validate points and patterns in 4D (3D cine) PC-MRI blood-flow data. PC-MRI blood-flow data consists of 3D vector fields for several time points within one heart cycle. The data of each time point is combined data from measurements during several heart cycles. We investigate three methods to obtain patterns in the blood-flow field. The winding number method was adapted to make it applicable to vector fields[2,3], and to extract critical points within the blood flow. Furthermore, we have inspected the ¿2 criterion by Jeong and Hussein[4] for the detection of vortices. For flow-pattern recognition, a method proposed by Heiberg et al.[5] is examined. Although these methods provide intrinsically different results, they can be compared because they all indicate the presence of patterns in the blood-flow field. The robustness of the winding number algorithm and the ¿2 criterion are investigated on several artificial data sets under increasing noise levels. The winding number is able to extract saddle points in images with a signal to noise ratio (SNR) of 4 or larger. The ¿2 criterion is able to extract vortices in images with a SNR of 8 or larger. This difference is largely due to the fact that the winding number uses more information from the local environment. In future work, robustness experiments regarding the pattern matching method will be performed. We expect the pattern matching method to be more robust than the winding number and ¿2 criterion approaches, because no derivatives are used, and more information from the local environment is incorporated. Furthermore, an advantage of the pattern matching method is that it can detect multiple patterns. Initial experiments on artificial data sets show promising results. This will be further investigated using measured blood-flow velocity data