EchoGLAD: Hierarchical Graph Neural Networks for Left Ventricle Landmark Detection on Echocardiograms
Authors: Masoud Mokhtari, Mobina Mahdavi, Hooman Vaseli, Christina Luong, Purang Abolmaesumi, Teresa S. M. Tsang, Renjie Liao
Abstract: The purposeful analysis of the left ventricle chamber of the middle requires detecting 4 landmark areas and measuring the inside dimension of the left ventricle and the approximate mass of the surrounding muscle. The vital factor downside of automating this course of with machine finding out is the sparsity of scientific labels, i.e., just some landmark pixels in a high-dimensional image are annotated, fundamental many prior works to carefully rely on isotropic label smoothing. However, such a label smoothing method ignores the anatomical knowledge of the image and induces some bias. To cope with this downside, we introduce an echocardiogram-based, hierarchical graph neural group (GNN) for left ventricle landmark detection (EchoGLAD). Our principal contributions are: 1) a hierarchical graph illustration finding out framework for multi-resolution landmark detection by means of GNNs; 2) induced hierarchical supervision at completely totally different ranges of granularity using a multi-level loss. We contemplate our model on a public and a private dataset beneath the in-distribution (ID) and out-of-distribution (OOD) settings. For the ID setting, we get hold of the state-of-the-art indicate absolute errors (MAEs) of 1.46 mm and 1.86 mm on the two datasets. Our model moreover displays greater OOD generalization than prior works with a testing MAE of 4.3 mm
