Hierarchical Convolution-Based Multilayer Perception for Denoising 3D MRI to Enhance Diagnostic Confidence in...

Cerebral small vessel disease (CSVD) manifestations in magnetic resonance (MR) images play a pivotal role as essential indicators for accurate diagnosis. However, the presence of noise in MR images significantly degrades image quality, thus compromising the precision of lesion detection and disease diagnosis. Although deep learning with residual architectures has shown promise in MR denoising tasks, current methods face several challenges, such as issues with model convergence, limited generalization capabilities, and oversmoothing, all of which collectively hinder denoising performance. Our objective is to enhance denoising performance by introducing a new model named the hierarchical convolution-based multi-layer perceptron (HC-MLP), specifically designed to improve the diagnostic confidence of CSVD. Our HC-MLP framework comprises three primary components: 1) The inclusion of MLP modules mitigates bias caused by pure CNN models. 2) The straightforward structures of MLPs and convolutional neural networks (CNNs) simplify training and improve generalization. 3) The use of voxel-wise input and the integration of the residual MLP structure partially address the oversmoothing issue. Extensive experiments have been conducted on two public datasets (UK Biobank and the anatomical tracings of lesion after stroker (ATLAS)) and an external dataset to test the effectiveness of HC-MLP. A total of 120 brain MRI scans from the UK Biobank and 120 brain MRI scans from ATLAS with CSVD were randomly chosen for model training and testing. For external testing, all 29 subjects with various MRI features of CSVD were included. Peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and normalized mean squared error (NMSE) were used for model evaluation. The Mann-Whitney-Wilcoxon two-sided test was used for score comparisons. Furthermore, two senior radiologists scored the results of the denoising performance. The experimental results show that HC-MLP significantly outperforms several state-of-the-art denoising algorithms, achieving a substantial improvement in PSNR (6.91% increase on UK Biobank and 5.31% on ATLAS) and SSIM (3.67% increase on UK Biobank and 2.27% on ATLAS). The CSVD recovery results further illustrate the superior performance of HC-MLP. Moreover, the performance of HC-MLP has been confirmed by radiologists. The proposed HC-MLP not only achieves significant enhancements in denoising 3D MR images but also successfully restores key features of CSVD that may be compromised by simulated noise in MR images, thereby improving the diagnostic confidence of CSVD.

The online version contains supplementary material available at 10.1007/s43657-025-00219-8.