Artificial intelligence-powered four-fold upscaling of human brain synthetic metabolite maps.

Abstract

ObjectiveCompared with anatomical magnetic resonance imaging modalities, metabolite images from magnetic resonance spectroscopic imaging often suffer from low quality and detail due to their larger voxel sizes. Conventional interpolation techniques aim to enhance these low-resolution images; however, they frequently struggle with issues such as edge preservation, blurring, and input quality limitations. This study explores an artificial intelligence-driven approach to improve the quality of synthetically generated metabolite maps.MethodsUsing an open-access database of 450 participants, we trained and tested a model on 350 participants, evaluating its performance against spline and nearest-neighbor interpolation methods. Metrics such as structural similarity index, peak signal-to-noise ratio, and learned perceptual image patch similarity were used for comparison.ResultsOur model not only increased spatial resolution but also preserved critical image details, outperforming traditional interpolation methods in both image fidelity and edge preservation.ConclusionsThis artificial intelligence-powered super-resolution technique could substantially enhance magnetic resonance spectroscopic imaging quality, aiding in more accurate neurological assessments.