Bidirectional Mapping Perception-enhanced Cycle-consistent Generative Adversarial Network for Super-resolution of Brain MRI images

As an effective tool for visualizing neurodegeneration, high-resolution structural magnetism facilitates quantitative image analysis and clinical applications. Super-resolution reconstruction technology allows to improve the resolution of images without upgrading the scanning hardware. However, existing super-resolution techniques relied on paired image data sets and lacked further quantitative analysis of the generated images. In this study, we proposed a semi-supervised generative adversarial network (GAN) model for super-resolution of brain MRI, and the synthetic images were evaluated using various quantitative measures. This model adopted the cycle-consistency structure to allow for a mixture of unpaired data for training. Perceptual loss was further introduced into the model to preserve detailed texture features at high frequencies. 363 subjects with both high-resolution (HR) and low-resolution (LR) scans and 217 subjects with HR scans only were used for model derivation, training, and validation. We extracted multiple voxel-based and surface-based morphological features of the synthetic and real 3D HR images for comparison. We further evaluated the synthetic images in the differential diagnosis of diseases. Our model achieved superior mean absolute error (0.049 +/- 0.021), mean squared error (0.0059 +/- 0.0043), peak signal-to-noise ratio (29.41 +/- 3.71), structural similarity index measure (0.914 +/- 0.048). Eight morphological metrics, both voxel-based and surface-based, showed significant agreement (P<0.0001). The gap of accuracy in disease diagnosis between synthetic and real HR images was within 5% and significantly outperformed the LR images. Our proposed model enables the reconstruction of HR MRI and could be used accurately for image quantification.