Feature Of The Week 1/8/12: Ultrafast Processing and Rendering of OCT Data using a GPU
Processing of the large volumetric data is the bottleneck of the current FD-OCT systems. Solving this issue enables real time visualisation. Various groups have applied utilising the full power of GPU for OCT data processing allows to realise the full potential of OCT. GPU based data processing to OCT. Here we present a GPU-CPU based processing methodology as well evaluates the performance of various GPUs.
In standard processing, interference data are copied to GPU and entire computation is carried out on GPU. Processed data are stored in GPU memory, which allows rendering and visualising the data.
To benchmark the processing rate of different GPUs, OCT data acquired using a swept-source OCT (EX1301, Michelson Diagnostic) is processing flowing standard processing protocols. Results shows processing rate of over 1.8 millions A-scans can be processed using one of the high end GPU. On the other hand processing rate of nearly 100,000 A-scans is achieved using a low end GPU.
Furthermore, to verify the numerical accuracy of the results, same data is processing on CPU. Absolute difference between the GPU and CPU is calculated. Taking the CPU as the reference results, any difference in the results greater than a given tolerance is taken as the fail. And number of failed were counted. The study suggests that the GPU based calculations are comparable to the CPU based values. Most of the error is dues to difference numerical precision of the hardware as well as the difference in the accuracy of the libraries used.
An easy to implement library (DLL) of standard OCT processing and post processing functions are available on request. The standard processing includes different interpolation methods. Currently we are developing other GPU based post processing methods.
For more information see recent Article. Courtesy Janarthanan Rasakanthan.