Hybrid CUDA acceleration of wavelet-based MRI image denoising

Background Magnetic resonance imaging is a key modality in medical diagnostics, but its quality is often degraded by noise, which reduces image clarity and may affect subsequent analysis. Wavelet-based denoising is effective for preserving anatomical structures, but its practical use is limited by computational cost. This study aimed to develop a hybrid graphics processing unit execution method for wavelet-based denoising of two-dimensional magnetic resonance images with high throughput and stable reconstruction quality. Methods A hybrid CUDA-based implementation of wavelet denoising was developed using a standard discrete wavelet transform with adaptive soft-thresholding as the denoising core. The proposed execution method combines data-level and task-level parallelism within a unified graphics processing unit pipeline and includes dual-stream vertical-pass scheduling, fused reconstruction kernels, constant-memory filter taps, symmetric boundary handling, and selective loop unrolling. Experiments were performed on 3,064 grayscale brain magnetic resonance images of size ×512 512 using synthetic Gaussian noise and simulated Rician noise and motion artifacts. An additional sensitivity analysis was conducted to assess the influence of wavelet basis, filter length, and threshold scaling on PSNR, SSIM, mean squared error, and runtime. Reconstruction quality was evaluated using peak signal-to-noise ratio, structural similarity index, and mean squared error, while computational performance was assessed using runtime and throughput. Results were summarized over repeated runs using mean values, standard deviations, and 95% confidence intervals. Results Under Gaussian noise, the proposed method achieved peak signal-to-noise ratio values of approximately 29–30 dB and structural similarity values around 0.90 while maintaining a throughput of 215–220 images per second. Compared with a reproduced data-parallel graphics processing unit wavelet baseline, the proposed implementation reduced runtime by 20%–30%. For Rician noise and motion-related artifacts, the method maintained stable slice-level throughput and consistent reconstruction behavior across the evaluated degradation types. Conclusion The proposed hybrid CUDA execution method provides an effective high-throughput solution for wavelet-based denoising of two-dimensional magnetic resonance images. By combining dual-stream scheduling, fused reconstruction, and memory-aware execution, the method improves runtime performance while preserving reconstruction quality.