RAPID 3D TRACING OF THE MOUSE BRAIN NEUROVASCULATURE WITH LOCAL MAXIMUM INTENSITY PROJECTION AND MOVING WINDOWS

Neurovascular models have played an important role in understanding neuronal function or medical conditions. In the past few decades, only small volumes of neurovascular data have been available. However, huge data sets are becoming available with high throughput instruments like the Knife-Edge Scanning Microscope (KESM). Therefore, fast and robust tracing methods become necessary for tracing such large data sets. However, most tracing methods are not effective in handling complex structures such as branches. Some methods can solve this issue, but they are not computationally efficient (i.e., slow). Motivated by the issue of speed and robustness, I introduce an effective and efficient fiber tracing algorithm for 2D and 3D data. In 2D tracing, I have implemented a Moving Window (MW) method which leads to a mathematical simplification and noise robustness in determining the trace direction. Moreover, it provides enhanced handling of branch points. During tracing, a Cubic Tangential Trace Spline (CTTS) is used as an accurate and fast nonlinear interpolation approach. For 3D tracing, I have designed a method based on local maximum intensity projection (MIP). MIP can utilize any existing 2D tracing algorithms for use in 3D tracing. It can also significantly reduce the search space. However, most neurovascular data are too complex to directly use MIP on a large scale. Therefore, we use MIP within a limited cube to get unambiguous projections, and repeat the MIP-based approach over the entire data set. For processing large amounts of data, we have to automate the tracing algorithms. Since the automated algorithms may not be 100 percent correct, validation is needed. I validated my approach by comparing the traced results to human labeled ground truth showing that the result of my approach is very similar to the ground truth. However, this validation is limited to small-scale real-world data due to the limitation of the manual labeling. Therefore, for large-scale data, I validated my approach using a model-based generator. The result suggests that my approach can also be used for large-scale real-world data. The main contributions of this research are as follows. My 2D tracing algorithm is fast enough to analyze, with linear processing time based on fiber length, large volumes of biological data and is good at handling branches. The new local MIP approach for 3D tracing provides significant performance improvement and it allows the reuse of any existing 2D tracing methods. The model-based generator enables tracing algorithms to be validated for large-scale real-world data. My approach is widely applicable for rapid and accurate tracing of large amounts of biomedical data.