SUPER-RESOLUTION OF DMRI BASED ON DENSENET WITH FA GUIDANCE

High-resolution (HR) imaging is essential for exploring the microstructure in the brain and thus plays a crucial role in facilitating early and accurate medical diagnosis. However, it often reduces the signal-to-noise ratio and limits coverage area due to long scanning times and increased sensitivity to device variability. To address this, we developed a new neural network architecture called Fractional Anisotropy (FA)-guided Densely Connected Super-Resolution network (FADCSR). We used a patch-based approach and advanced convolutional networks with adaptive three-dimensional (3D) dense connectivity to efficiently process the FA features, which are vital for guiding the super-resolution reconstruction of diffusion-weighted images. This enhanced the quality of the reconstructed HR images and preserved the microscopic structural features. We evaluated the FADCSR model against other leading super-resolution methods using the Human Connectome Project datasets. Our results showed superior super-resolution performance of FADCSR in reconstructing 3D diffusion-weighted magnetic resonance imaging, demonstrating its effectiveness and potential for broader applications in medical imaging technology.