Hybrid CUDA acceleration pushes wavelet-based MRI denoising past 200 images per second

Magnetic resonance imaging produces some of the most diagnostically valuable pictures in modern medicine, but every scan carries an unwanted passenger: noise. Thermal fluctuations in the electronics, the statistical quirks of the magnitude reconstruction process, and the involuntary movements of patients all conspire to blur fine anatomical detail and undermine the automated analysis systems increasingly used in clinical workflows. A research team from Lviv Polytechnic National University and West Ukrainian National University in Ukraine has now reported a GPU execution strategy that attacks the computational cost of one of the most respected classical denoising methods, wavelet filtering, and pushes it into near-real-time territory. Writing in Frontiers in Signal Processing, Lesia Hentosh, Yana Vyklyuk and Oleh Berezsky describe a hybrid CUDA implementation that cleans 2D brain MRI slices at roughly 215 to 220 images per second while keeping reconstruction quality essentially unchanged.

The choice of wavelet denoising as the target is deliberate. Deep learning denoisers, including convolutional networks and CNN–Transformer hybrids, can achieve impressive reconstruction metrics, but they depend on large representative training sets, careful adaptation to scanner protocols, and hardware-specific inference behavior. In settings where paired noisy and clean acquisitions are unavailable, or where the denoising stage must remain transparent and auditable, a deterministic alternative has clear appeal. The discrete wavelet transform requires no training, separates low-frequency anatomical structure from high-frequency noise through a mathematically explicit multiresolution decomposition, and offers a well-defined reconstruction pathway. Its weakness has always been speed: decomposition, thresholding, and reconstruction executed sequentially impose a computational burden that has limited clinical deployment, particularly in latency-sensitive pipelines.

The denoising core of the new method is deliberately classical. Each 512 × 512 grayscale brain MRI slice is decomposed with a single-level two-dimensional discrete wavelet transform using biorthogonal 4.4 filters, a basis chosen for its near-symmetric, linear-phase reconstruction properties that preserve anatomical edge geometry better than asymmetric orthogonal wavelets such as Daubechies, and more smoothly than the block-prone Haar wavelet. The transform splits the image into four subbands: a low-frequency approximation and three high-frequency detail components capturing horizontal, vertical, and diagonal information. Soft thresholding is then applied adaptively to the detail subbands, with thresholds derived from the median absolute deviation of the diagonal subband and scaled by subband-specific weights, so that the diagonal component receives stronger suppression while oriented edges are treated more conservatively. The inverse transform then reconstructs the cleaned image.

What distinguishes the study is not this mathematical machinery, which follows standard wavelet denoising formulations, but the execution-level reorganization on the graphics processing unit. In a conventional data-parallel CUDA realization, decomposition, thresholding, and reconstruction run as consecutive kernels with synchronization between stages, and intermediate results repeatedly travel to and from global memory. The Ukrainian team instead treats the pipeline as a dependency graph. After horizontal filtering, the low-frequency and high-frequency branches are independent, so they are assigned to two separate CUDA streams during the vertical pass. The same dual-stream principle governs inverse vertical reconstruction, where the approximation-plus-horizontal-detail pair and the vertical-plus-diagonal pair are processed concurrently. The final horizontal reconstruction is performed by a fused kernel that combines filtering and accumulation in a single launch, avoiding an additional write-read cycle through global memory.

Further memory-aware refinements complete the design. The short bior4.4 filter coefficients are stored in constant GPU memory, allowing broadcast access by all threads with minimal latency, and symmetric boundary handling is performed directly inside the kernels rather than by constructing padded buffers. Selective loop unrolling reduces iteration counts in the vertical filtering kernels. The authors report that the effective latency of the overlapped stages is governed by the slower branch plus synchronization overhead, rather than the sum of both branch runtimes, which is the mathematical source of the speedup. Profiling on an NVIDIA Tesla T4 showed that the dual-stream vertical filtering and inverse vertical reconstruction stages dominate kernel time, accounting for 24.4 and 23.2 percent respectively, with achieved occupancy between 68 and 86 percent across kernel groups.

The experimental evaluation covered all 3,064 grayscale brain MRI slices from a publicly available brain tumor classification dataset, converted to single-channel 512 × 512 float32 images. Three controlled degradation models were applied: additive Gaussian noise with standard deviations from 5 to 25, simulated Rician noise generated by adding independent Gaussian components to the real and imaginary channels of a complex signal before magnitude reconstruction, and simplified motion artifacts produced by rigid in-plane translations of up to 15 pixels and rotations of up to 10 degrees. Under Gaussian noise, the method achieved a peak signal-to-noise ratio of approximately 30.63 dB and a structural similarity index of 0.922, while Rician noise yielded 28.20 dB and 0.845, and motion artifacts 26.51 dB and 0.880, with narrow 95 percent confidence intervals across repeated runs.

The computational results are striking. Compared with a CPU implementation, the GPU version achieved a maximum acceleration factor of 124 times for a single image, stabilizing at roughly 68 to 70 times for larger batches, with per-image runtime remaining nearly constant at about 4.4 to 4.6 milliseconds regardless of dataset size. Crucially, the acceleration did not alter the reconstruction: PSNR and SSIM values for CPU and GPU implementations were practically identical, and a difference heatmap confirmed that CPU–GPU discrepancies are confined to negligible floating-point deviations. Against a reproduced data-parallel GPU baseline executed under identical conditions, the hybrid method reduced end-to-end per-image runtime from 5.96 to 4.65 milliseconds on the full dataset, a gain of approximately 22 percent, with batch-level improvements ranging from about 4 to nearly 25 percent.

A sensitivity analysis reinforced the robustness of the chosen configuration. Comparing Haar, Daubechies-4, Symlet-4, Coiflet-2, and biorthogonal 4.4 bases, the authors found that Haar offered the lowest runtime but the weakest structural preservation, while smoother, longer filters improved reconstruction metrics at moderate runtime cost. The bior4.4 basis delivered the best balance, achieving the highest PSNR and SSIM while sustaining throughput above 200 images per second. Varying the global threshold scaling coefficient from 0.6 to 1.4 changed reconstruction metrics systematically but left runtime essentially unchanged, confirming that the selected parameters reflect a stable operating point rather than a tuned artifact of the benchmark.

The authors are careful to frame the contribution honestly. The evaluation was a controlled computational benchmark on 2D slices with synthetic and simulated degradations, not a clinical validation study. Real in vivo MRI contains spatially varying noise from coil sensitivity profiles, field inhomogeneities, parallel imaging reconstruction, and patient-dependent motion patterns that the simplified models do not fully reproduce. The method also does not yet exploit inter-slice correlations in volumetric data; a true 3D extension would require a separable three-dimensional wavelet transform producing eight subbands, redesigned stream scheduling along three axes, and memory-aware volume tiling, which the team identifies as a dedicated direction for future work. Multi-architecture GPU benchmarking likewise remains to be performed.

Within those limits, the study offers a compelling demonstration that classical, interpretable image processing can be made fast enough for modern clinical pipelines through intelligent parallel execution rather than model substitution. A deterministic, training-free denoiser running at more than 200 slices per second could serve as a lightweight preprocessing stage before segmentation, classification, or quantitative analysis modules, reducing waiting time in batch workflows without becoming the dominant bottleneck. As artificial intelligence continues to reshape medical imaging, this work suggests that the older mathematical tools, when reorganized for the hardware they actually run on, still have a great deal to contribute.

Subject of Research: GPU-accelerated wavelet denoising of 2D brain MRI images using hybrid CUDA parallelization

Article Title: Hybrid CUDA acceleration of wavelet-based MRI image denoising

Article References: Hentosh, L., Vyklyuk, Y., & Berezsky, O. (2026). Hybrid CUDA acceleration of wavelet-based MRI image denoising. Frontiers in Signal Processing, 6, Article 1862897. https://doi.org/10.3389/frsip.2026.1862897

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1862897

Keywords: MRI denoising, CUDA, GPU computing, wavelet transform, medical imaging, parallel processing, image reconstruction, signal processing, Hybrid, acceleration, wavelet-based, image