Generation of 3D Super-resolution FLAIR Images from Heterogeneous 2D FLAIR Acquisitions in Epilepsy Patients

Motivation: 3D FLAIR often provides essential information for clinical MRI review, especially for detecting epileptic lesions. 2D FLAIR is still commonly used in many healthcare settings. Goal(s): We aim to develop a framework for synthesizing high-resolution 3D FLAIR images from heterogenous 2D axial FLAIR acquisitions. Approach: We used paired 2D and 3D FLAIR images from 31 healthy controls and 17 epilepsy patients to train a generative-adversarial network (GAN). Structural similarity index (SSIM) between the generated images and the original 3D FLAIR images was calculated. Results: The average SSIM was 0.898±0.020 for our generated 3D FLAIR images, significantly higher than B-spline interpolation (0.845±0.021, p=0.015). Impact: Our method can assist clinicians to evaluate the existence and the extent of epileptic lesions in data-limited situations in various health care settings. The generated images may also serve as inputs to inform AI models for automated lesion detection.