Invertible Sharpening Network for MRI Reconstruction Enhancement

High-quality MRI reconstruction plays a critical role in clinical applications. Deep learning-based methods have achieved promising results on reconstruction. However, most state-of-the-art were designed to optimize the evaluation metrics commonly used for natural images, such as PSNR and SSIM, whereas visual quality is not primarily pursued. Compared fully-sampled reconstructed images are often blurry, where high-frequency features might be sharp enough confident diagnosis. To this end, we propose an invertible sharpening network (InvSharpNet) improve of reconstructions. During training, unlike traditional that learn map input data ground truth, InvSharpNet adapts backward training strategy learns blurring transform from truth (fully-sampled image) (blurry reconstruction). inference, learned can inverted leveraging network's invertibility. The experiments various datasets demonstrate sharpness with few artifacts. also evaluated by radiologists, indicating better diagnostic confidence our proposed method.