PT  - JOURNAL ARTICLE
AU  - Leyun Pan
AU  - Caixia Cheng
AU  - Antonia Dimitrakopoulou-Strauss
AU  - Uwe Haberkorn
AU  - Ludwig Strauss
TI  - MLPI-Grid: A grid for CPU-intensive and machine-learning based PET parametric imaging computing
DP  - 2011 May 01
TA  - Journal of Nuclear Medicine
PG  - 1988--1988
VI  - 52
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/52/supplement_1/1988.short
4100  - http://jnm.snmjournals.org/content/52/supplement_1/1988.full
SO  - J Nucl Med2011 May 01; 52
AB  - 1988 Objectives The machine-learning (ML) based parametric imaging (PI) is a robust and reproducible solution to quantitative analysis of dynamic PET data instead of iterative fitting method. Unfortunately, it is a cpu-intensive method. The study established a computing grid, which divides the parametric imaging work to several voxel-based tasks that can be parallel calculated by agents on the grid. Methods The MLPI-Grid is a java application that can be deployed on a cross-platform environment. The three-tiers architecture including clients, controllers, and agents work together to streamline the PI process. A client loads the PET data (the imaging and kinetic model) and creates the parametric imaging jobs according to the number of voxels and agents. Then the client submits jobs to the controller. The controller queues tasks, distributes those tasks to agents, and handles task reassignment. Each agent runs the machine learning algorithm to produce the parametric image. It fully utilizes a historical reference database to finds a balance between fitting to history data and unseen target curve, which can build a moderate kinetic model to directly deal with noisy PET data. The controller collects job results that exist as a map from a voxel coordinates to a set of model parameters and assembles them to the parametric images. Results The grid was installed and evaluated on a local network including 3 12-Cores and 10 8-Cores Mac Pro systems(Apple Computer Inc., Cupertino, CA, USA). Basically, the performance is in direct proportion with the number and the performance of agents on the grid. For example, the parametric images of a 512*512*56*28 dynamic PET data can be calculated in less than 1 hour instead of 12 hours. Conclusions The parallel feature of parametric imaging makes itself natural for grid computation. Our MLPI-Grid fulfilled the speedup demand of image processing and can also be expanded to other cpu-intensive tasks like image fusion