Artificial intelligence with spatial-attention for ventricular volumes and mass assessment in cardiac magnetic resonance images with susceptibility artifacts

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Introduction Segmentation of cardiac magnetic resonance (CMR) images is an essential step for evaluating dimensional and functional ventricular parameters but may limited by noise artefacts, such as the increasing diffusion of implantable devices causing distortion of the magnetic field. Purpose The aim of this paper is to investigate the accuracy of a deep learning (DL) approach for automatic segmentation of cardiac structures from CMR images characterized by magnetic susceptibility artifact images. Methods A retrospective database of SAX (short axis) cine CMR images obtained from 216 patients using a 1.5 Tesla scanner was used to develop and validate the proposed DL approach, subdividing patients into two subgroups, considering the presence (n=86) or absence (n=130) of a cardiac implantable electronic device. A novel convolutional neural network (CNN) was proposed to extract the left (LV) and right (RV) ventricle endocardium and left ventricle epicardium (LV mass). The proposed network takes advantage from a spatial attention module which introduces an adaptive feature refinement through a spatial weighting strategy to identify salient image regions even in the presence of artifacts. To improve segmentation, especially for images with artifacts, a compounded loss function is introduced. Segmentation results were assessed against manual tracings and commercial CMR analysis software. Results Results of correlation and Bland-Altman analyses on images with artifacts were reported for the proposed CNN and the commercial software (CS) in respect to the manual gold standard (GT). Also interobserver variability between O1 and O2 was tested. The proposed method reached higher segmentation accuracy than CS, with performance comparable to expert inter-observer variability (Picture 1). It is possible to appreciate how all the contours were properly depicted using the proposed CNN, with results comparable to the manual tracings. On the opposite, the CS resulted in imprecise segmentations of LV and RV endocardium and LV epicardium in presence of susceptibility artefacts caused by cardiac implantable electronic device affecting images' quality (Picture 2). Furthermore, the automatic approach was at least 100 times faster than manual segmentation in providing cardiac parameters. Conclusion The proposed method reached promising performance in cardiac segmentation from CMR images with susceptibility artifacts and alleviates time consuming expert physician contourn segmentation.