DFAN: Dual Frequency-Aware Network for 3D MRI volume super-resolution

Deep learning-based 3D MRI volume super-resolution (SR) is a key technology designed to reconstruct high-resolution MRI volumes from low-resolution counterparts for accurate diagnosis. However, existing 3D MRI volume super-resolution methods mainly adopt shallow 3D convolutional neural network (CNN) architectures with limited receptive fields, making it difficult to extract global and complex features. These methods also process entire feature maps without targeted extraction of high-frequency information, which may lead to poor reconstruction of edges and textures. To address these problems, we propose a novel Dual-Frequency Aware Network (DFAN) that combines the advantages of CNN and Transformer to process deep features of different frequencies. We utilize normal convolution and fast Fourier convolution in parallel to capture low-frequency feature globally. Additionally, we extract multi-directional gradient features within Transformer as high-frequency priors, which guide the attention mechanism to effectively aggregate high-frequency information. Cross-attention is employed to enhance the interaction between low-frequency and high-frequency features. To reduce the computation of Transformer, we design feature embedding and unembedding operations based on discrete cosine transform. Experiments on two public brain MRI datasets, IXI and BraTS 2021, demonstrate that DFAN surpasses state-of-the-art 3D MRI SR methods. On the IXI dataset, DFAN achieves PSNR/SSIM of 30.24/0.9075 at × 2, and 25.57/0.7480 at × 4. On the BraTS 2021 dataset, DFAN achieves 35.72/0.9638 at × 2, and 31.44/0.9115 at × 4. Segmentation experiments on the reconstructed BraTS 2021 dataset also indicate the potential of DFAN for clinical applications. • Employ various strategies to extract feature from different frequency domains. • Use gradient features as prior knowledge to aggregate high-frequency information. • Complement low-frequency and high-frequency information based on cross-attention. • Utilize the Discrete Cosine Transform to preserve global relationships of features. • Outperform other SOTA methods on two public MRI datasets in reconstruction quality.