Classifying MRI motion severity using a stacked ensemble approach.

Motion artifacts are a common occurrence in Magnetic Resonance Imaging exam. Motion during acquisition has a profound impact on workflow efficiency, often requiring a repeat of sequences. Furthermore, motion artifacts may escape notice by technologists, only to be revealed at the time of reading by the radiologists, affecting their diagnostic quality. There is a paucity of clinical tools to identify and quantitatively assess the severity of motion artifacts in MRI. An image with subtle motion may still have diagnostic value, while severe motion may be uninterpretable by radiologists and requires the exam to be repeated. Therefore, a tool for the automatic identification of motion artifacts would aid in maintaining diagnostic quality, while potentially driving workflow efficiencies. Here we aim to quantify the severity of motion artifacts from MRI images using deep learning. Impact of subject movement parameters like displacement and rotation on image quality is also studied. A state-of-the-art, stacked ensemble model was developed to classify motion artifacts into five levels (no motion, slight, mild, moderate and severe) in brain scans. The stacked ensemble model is able to robustly predict rigid-body motion severity across different acquisition parameters, including T1-weighted and T2-weighted slices acquired in different anatomical planes. The ensemble model with XGBoost metalearner achieves 91.6% accuracy, 94.8% area under the curve, 90% Cohen's Kappa, and is observed to be more accurate and robust than the individual base learners.