부산대학교 도서관

Time sequences of 3D images of cerebral and other vasculature blood flow during surgery and other medical procedures allow enhanced visual feedback. The visual feedback constitutes an enhancement over the existing 2D time series of Xray projections as it facilitates the detection and observation of pathological abnormalities such as stenoses, aneurysms, and blood clots. An algorithm that outputs 3D+T sequences by fusing a single static 3D model of the vasculature with two time sequences of 2D projections was presented in [1]. Practical clinical use demands that the reconstruction be completed within a 5 minute time frame. When compared to CPU implementations, past GPU-based CT (computational tomography) implementations typically achieved one order-of-magnitude speed improvement, still insufficient speed for this application. To obtain further needed GPU speedup, we exploit the sparse structure of blood vasculature in order to achieve a total of two orders-of-magnitude performance increase. Our GPU implementation generates a 3D+T time series reconstruction in 2 minutes, enabling real time clinical use and safer, shorter procedures. Included in our approach is an architecture-aware partitioning method that accelerates the solution to a wide class of variational problems.