Aims: With advances in CMOS sensor technology, optical mapping measurements can be performed at megapixels resolutions >1 Mpixel and frame rates >500 fps at staggering data rates >1 GB/s, two orders of magnitude than a decade ago utilizing CCD sensors. Removal of baseline wandering involves non-linear methods of >O(n^2) complexity and this study aims to present a GPU accelerated parallel image processing method suitable for near real-time removal of optical mapping baseline wandering.
Methods: Optical mappings were performed on isolated rabbit hearts with Di-4-ANBDQPQ Vm dye at resolution of 1600x1000 pixels at 500 fps. Baseline wandering removal was assessed with Keiser Window and Equiripple finite impulse response (FIR) filters in MATLAB with CPU processing (Intel i7, 8th generation) and GPU processing (NVIDIA 1080Ti 3584 CUDA cores), and native CUDA GPU processing, all in dual precision. Filter parameters are assessed with different slopes between stop band frequency (Fs) and pass band frequency (Fp), with Fs/Fp ratios ranging from 0.1 to 0.9.
Results: FIR filters processing time was timed for different Fs/Fp ratios exhibiting exponential complexity as the ratio approaches to 1. GPU acceleration in MATLAB resulted in processing speed increase of up to 10 times, while native CUDA GPU processing resulted in speed increase of up to 50 times compared to single core CPU processing. CUDA processing data flow rates ranged from 2.5 GB/s for Fs/Fp = 0.1 down to 300 MB/s for Fs/Fp = 0.9.
Conclusions: GPU gaming class cards nowadays have few thousands cores at affordable prices and are suitable tool to leverage technological progress in optical sensors developments with increased resolution/frames rates as image processing is suitable for parallelization. Near-real time processing capabilities allow optical mapping method to become the research tool for real-time experimental setups enabling on-site decision making process, and disrupting traditional off-line data analysis.