Remedios, Samuel W.; Wei, Shuwen.; Han, Shuo.; Zhang, Jinwei.; Carass, Aaron.; Schilling, Kurt G.; Pham, Dzung L.; Prince, Jerry L.; Dewey, Blake E. (2026).听.听Journal of Medical Imaging, 13(2), 024001.听
Magnetic resonance imaging, or MRI, is often collected as a stack of 2D slices because that can make scans faster and improve image quality for clinical use. But when software tries to analyze these scans as if they were full 3D images, it can struggle, especially when the slices are thick or have gaps between them. To address this, the researchers developed ECLARE, a new method that improves the resolution of these slice-based MRI scans without needing outside training data. ECLARE first estimates the shape of each slice鈥檚 signal, then learns from the image itself how to turn lower-resolution parts into higher-resolution ones, while also correcting for blur and making sure the image is resampled in a way that respects the original field of view. The method was tested on brain MRI data, including healthy T1-weighted scans and T2-FLAIR scans from people with multiple sclerosis, and compared with several existing image-enhancement methods. Across scans with slice thicknesses up to 5 mm and gaps up to 1.5 mm, ECLARE produced more accurate and visually similar images than the alternatives, including in important brain regions such as the ventricles, caudate, and white matter. Overall, the study suggests that ECLARE can make thick-slice MRI images more useful for 3D analysis, which could improve downstream tools that rely on detailed brain structure.
Fig.听1
Flowchart of our proposed method. The anisotropic input volume is fed independently into each of the three steps. First, in panel a (Sec.听), we estimate the slice excitation profile with ESPRESO.聽Second, in panel b (Sec.听), we extract HR in-plane 2D patches and use the PSF estimated from panel a to create paired training data. This training data are used to train the network聽饾憮饾渻聽with supervised learning. Third, in panel c (Sec.听), we extract LR through-plane 2D slices and superresolve them with the trained network聽饾憮饾渻聽from panel b. The superresolved slices are stacked and averaged, yielding the superresolved output volume.