Addressing inefficiencies in brain magnetic resonance imaging workflows with artificial intelligence

Abstract

Brain magnetic resonance imaging (MRI) is currently constrained by long scan times, limiting patient throughput and accessibility. Artificial Intelligence (AI) models have shown promise to make scans faster, therefore increasing patient throughput. However, simply increasing scan volume creates a secondary bottleneck: the finite capacity of radiologists to interpret a growing influx of data. To address this dual challenge, we propose a framework that integrates longitudinal imaging data to achieve significantly shorter scan times than current state-of-the-art methods. To streamline the resulting workload, we propose to leverage vast adult normative datasets to develop brain MRI foundation models that learn robust representations by leveraging domain-informed pretext tasks in which the training labels are easy to obtain, such as voxel-level brain age and Fazekas score predictions. This model will serve as a robust platform to automate, expedite, and enhance the precision of brain MRI analysis, ultimately bridging the gap between rapid data acquisition and actionable clinical insights. In this paper, we show examples of longitudinal MRI reconstruction that achieve up to 20× acceleration, adaptation of reconstruction models trained on adults to infants, voxel-level brain age prediction, and Fazekas score prediction.