Denoising medical images is essential for improving diagnostic precision in CT and MRI scans. This paper introduces a flexible hybrid framework that integrates multiscale decomposition methods (Empirical Mode Decomposition, Variational Mode Decomposition, Bidimensional EMD, and Multivariate EMD) with curvelet transform thresholding and conventional spatial filters. The approach was evaluated on a phantom dataset with controlled Rician noise, clinical CT images reconstructed using sharp (B50f) and medium (B46f) kernels, and MRI data acquired at different GRAPPA acceleration levels.
In phantom experiments, MEMD–Curvelet achieved the top SSIM (0.964) and PSNR (28.35 dB), alongside strong perceptual quality (NIQE ~7.55, BRISQUE ~38.8). For CT images, VMD–Curvelet and MEMD–Curvelet surpassed traditional filters, delivering SSIM >0.95 and PSNR >28 dB, particularly with sharp kernels. In MRI data, MEMD–Curvelet and BEMD–Curvelet minimized perceptual artifacts, cutting NIQE by up to 15% and BRISQUE by 20% versus Gaussian and median filters.
Comparisons with deep learning methods confirmed the framework’s strength: BM3D offered high fidelity but took 6.65 s per slice, while DnCNN matched SSIM (0.958) in just 2.33 s. Overall, the framework demonstrates superior noise removal and detail retention across imaging modalities, outperforming standalone filters or pure transform techniques. Its adaptability and speed highlight its value for clinical use in resource-constrained imaging scenarios.
Autors: Mahmoud Nasr, Krzysztof Brzostowski, Rafał Obuchowicz, Adam Piórkowski
Keywords: medical image denoising; curvelet transform; EMD; VMD; BEMD; MEMD; BM3D; DnCNN; CT kernels; MRI GRAPPA; rician noise; image quality metrics