Fei Yang, Abstract Co-Author: Nothing to Disclose

Jie Tian PhD, Presenter: Nothing to Disclose

Most of the existing medical image processing systems still uses CPU dominated memory and computation model. This becomes a big limitation as parallel heterogeneous systems becoming mature and demonstrating significantly better performance. We have designed and implemented a fully GPU-accelerated software platform for medical image processing and visualization. Based on the new platform, several segmentation, registration and rendering algorithms have been implemented and tested in image based diagnosis and surgical planning. A runtime GPU code generation mechanism is applied to equip GPU programming with the same level of flexibility as CPU programming.

We have partners from 3 different hospitals using medical images for different purposes. We built different applications according to their requirements using both open-source platforms VTK/ITK and our new platform, implementing basically the same function sets. According to the feedback, overall, our new platform out-performs the open-source toolkits by 52% in speed.

Although both the open-source society and commercial parties have been trying hard to develop platforms for medical image processing, the computation capability of most recent hardware cannot be fully utilized due to the difficulty of incorporating the new programming interfaces with the existing platform architectures. We have discovered the key to this problem, and we have successfully developed a new platform with both high-performance and high-flexibility. Several segmentation, registration and visualization algorithms have been implemented, including level-set, and push-relabel based graph-cut for segmentation, registration optimizers for both collinear and basic deformable transforms such as B-SPLINE, as well as an OpenCL and GLSL based ray-casting engine capable of rendering multiple volumetric and geometric entities simultaneously in the same scene.  

We designed and implemented a fully GPU-accelerated software platform for medical image processing and visualization. Better performance has been achieved in comparison with existing open-source platforms during multiple case studies.

Yang, F, Tian, J, A Fully GPU-accelerated Platform for Medical Image Processing and Visualization.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14012465.html