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
Summary: The purposeful evaluation of the left ventricle chamber of the center requires detecting 4 landmark areas and measuring the inner dimension of the left ventricle and the approximate mass of the encompassing muscle. The important thing problem of automating this process with machine studying is the sparsity of scientific labels, i.e., only some landmark pixels in a high-dimensional picture are annotated, main many prior works to closely depend on isotropic label smoothing. Nevertheless, such a label smoothing technique ignores the anatomical data of the picture and induces some bias. To deal with this problem, we introduce an echocardiogram-based, hierarchical graph neural community (GNN) for left ventricle landmark detection (EchoGLAD). Our principal contributions are: 1) a hierarchical graph illustration studying framework for multi-resolution landmark detection through GNNs; 2) induced hierarchical supervision at totally different ranges of granularity utilizing a multi-level loss. We consider our mannequin on a public and a non-public dataset below the in-distribution (ID) and out-of-distribution (OOD) settings. For the ID setting, we obtain the state-of-the-art imply absolute errors (MAEs) of 1.46 mm and 1.86 mm on the 2 datasets. Our mannequin additionally exhibits higher OOD generalization than prior works with a testing MAE of 4.3 mm
