Unsupervised Standard Plane Synthesis in Population Cine MRI via Cycle-Consistent Adversarial Networks

In clinical studies or population imaging settings, cardiac magnetic resonance (CMR) images may suffer from artifacts due to variability in the breath-hold position adopted by the patient during the scan. Consistent orientation of image planes with respect to the cardiac ventricles in CMR sequences forms a crucial step in the assessment of cardiac function via parameters such as the Ejection Fraction (EF) and Cardiac Output (CO) of both ventricles, which are the most immediate indicators of normal/abnormal cardiac function. In this paper, we present a novel unsupervised approach for the realistic transformation of acquired CMR images to a standard orientation using Cycle-Consistent Adversarial Networks (Cycle-GANs). We tackle this challenge by splitting the problem into two principal subtasks. First, we consider a bidirectional generator mapping between the re-oriented image and the original, hence allowing direct comparison to the input image without the need to resort to paired training data. Second, we devise a novel loss function incorporating intensity and orientation terms, and aims to produce images of high perceptual quality. Extensive experiments conducted on the CMR images in the UK Biobank dataset demonstrate that the images rendered by our model can improve the accuracy of the image derived cardiac parameters.