Three sub-projects are described below: Fast Random Walker, Fast Parallel MRF, and Fast GPU Kinetic Models.
Fast Random Walker
Fast Random Walker, FastRW, speeds up the RW with prior segmentation algorithm through using additional pre-computation, please visit the new FastRW website for more details and code.
Fast Parallel MRF
Markov Random Fields (MRFs) are of great interest to the medical image analysis community but suffer from high computational complexity and difficulties in parameter selection. For these reasons, efforts have been made to develop more efficient algorithms for solving MRF optimization problems in order to enable reduced run-times and better interactivity. However, these algorithms are often implemented in serial and thus do not benefit from multi-core technology. We demonstrate a parallelized implementation of a popular MRF optimization algorithm, belief propagation, and use it to perform a binary image segmentation. By utilizing modern, lightweight parallel-programming techniques we are able to achieve a speedup of approximately 8 times, reducing the average segmentation time of a single 600×450 image from 12.7s to 1.6s.
Shane Mottishaw, Sergey Zhuravlev, Lisa Tang, Alexandra Fedorova, and Ghassan Hamarneh. An Evaluation of Parallelization Techniques for MRF Image Segmentation. In International Workshop on High-Performance Medical Image Computing for Image-Assisted Clinical Intervention and Decision-Making (MICCAI HP), pages 10 pages, 2010.
Fast GPU Kinetic Models
We address the problem of quickly and accurately fitting compartmental models to large numbers of measurements, with application to dynamic positron emission tomograph imaging. An analytic method of calculating the model output function is developed, and compared to the numerical method used by COMKAT, the de facto kinetic modeling software. A performance increase of over 13 times is demonstrated. Also, we describe a GPU implementation of the Levenberg–Marquardt optimization algorithm, using the developed analytic formulation, and apply it to compartmental model fitting. A fitting rate of 74,906 TAC/s was achieved, allowing a dPET volume to be fit in 13.8 seconds. In com- parison, COMKAT would take 21.5 days to perform the same fitting.
Ben Smith, Ghassan Hamarneh, and Ahmed Saad. Fast GPU Fitting of Kinetic Models for Dynamic PET. In International Workshop on High-Performance Medical Image Computing for Image-Assisted Clinical Intervention and Decision-Making (MICCAI HP), pages 10 pages, 2010.