A data-driven semantic segmentation model for direct cardiac functional analysis based on undersampled radial MR cine series.

Abstract

Imageacquisition and subsequent manual analysis of cardiac cine MRIis time-consuming. The purpose of this study was to train and evaluate a 3D artificial neural network for semanticsegmentation ofradially undersampledcardiacMRI to accelerate both scan time and postprocessing. A database ofCartesianshort-axisMR images of the heart(148,500 images, 484 examinations)was assembled fromanopenly accessibledatabaseandradial undersampling wassimulated.A3D U-Net architecture waspretrainedfor segmentation of undersampled spatiotemporal cineMRI.Transfer learning wasthen performed using samples from asecond database, comprising108non-Cartesian radial cine series of the midventricularmyocardium to optimize the performancefor authentic data.The performance was evaluated for different levelsof undersampling by theDice similarity coefficient(DSC)with respecttoreferencelabels, as well as by deriving ventricular volumes and myocardial masses. Without transfer learning,thepretrained modelperformedmoderately ontrueradial data [maximum number of projections tested, P=196;DSC =0.87(left ventricle),DSC=0.76(myocardium),and DSC =0.64(right ventricle)]. After transfer learning with authentic data, the predictions achieved human levelevenforhighundersamplingrates (P=33, DSC=0.95, 0.87, and 0.93)withoutsignificant difference compared withsegmentations derived fromfully sampled data. A 3D U-Net architecture can be used for semantic segmentation of radially undersampled cine acquisitions, achieving a performance comparable with human experts in fully sampled data. This approach can jointly acceleratetime-consumingcineimage acquisition andcumbersome manual image analysis.