Deep learning based processing for quantitative myocardial perfusion MRI

摘要

Quantitative myocardial perfusion cardiovascular magnetic resonance (CMR) provides a user-independent assessment of myocardial perfusion status for the non-invasive diagnosis of myocardial ischaemia, with high prognostic value [1]. However, it currently has limited use in clinical practice due to the challenging post-processing required, particularly for the segmentation of the images. We propose an automated method for processing the images prior to quantitative analysis based on deep learning techniques, and in particular convolutional neural networks (CNNs). Subjects and Methods: A sequence of object detection and image segmentation tasks is performed. There is first a classifier that identifies the time dynamic corresponding to peak enhancement in the left ventricle (LV). A CNN then predicts a bounding box around the LV and myocardium prior to motion correction. After motion correction [2], the myocardium is segmented and the right ventricular insertion point is detected. Due to the use of motion correction, segmentations and landmarks can be propagated to all time dynamics. These are then used in the tracer-kinetic modelling. Each step is assessed individually followed by a comparison of the automated and manually obtained myocardial blood flow (MBF) values on a segmental level. The full pipeline is shown in.