GPU Based Denoising Filter for Knee MRI

MRI is a popularly used technique for diagnosing muscle and skeletal disorders, especially of the knee. For accuracy in diagnosis, the rician noisy knee image needs to be filtered using efficient denoising algorithm. In recent years, the spatial neighborhood bilateral filter is being explored by researchers for its capacity to retain edges and tissue structures. It is noted that increase in image resolution slows down performance of the bilateral filter effectively discouraging its use. The research work proposes a cost-effective accelerated solution to the problem by implementing CUDA-based bilateral filter as applied to T2-weighted sagittal knee MRI slice. The work suggests use of GPU shared memory for optimized implementation and better speedup. The speedup achieved for 3.96 Mpixel knee MR image is 114.27 times more than that of its CPU counterpart. The results indicate average occupancy of 90.15% for image size of 630² pixels, indicating effective parallelization. Also, over varying rician noise levels, the average PSNR achieved is 21.83455 dB indicating good filter performance.