GHOST-CAT: New AI network generates 3D heart meshes from echo with 0.87 Dice

A team of researchers from the University of Auckland (Edward Ferdian, Debbie Zhao, Alistair A. Young, Martyn P. Nash) has released GHOST-CAT, a novel deep learning pipeline specifically designed for mesh generation from 3D echocardiograms. Echocardiography presents unique challenges—low contrast-to-noise ratio, conical field of view, and acoustic shadowing—that often trip up standard segmentation models. GHOST-CAT tackles these with a two-stage architecture: first, a convolutional neural network extracts features; then a graph convolutional network combined with a transformer refines the mesh to be topologically consistent and temporally coherent across the cardiac cycle.

The results on a held-out test set of 100 3D echo images are compelling. GHOST-CAT achieved a Dice coefficient of 0.87 ± 0.05 for the left ventricular cavity and 0.75 ± 0.07 for the myocardium, with mean surface distances of 3.3 ± 0.6 mm (endocardium) and 3.5 ± 0.5 mm (epicardium) against reference segmentations from cardiac MRI. These figures surpass current state-of-the-art methods. The reconstructed meshes allow automated computation of routine clinical indices like volume, mass, and strain, and can feed into biophysical digital twins for personalized medicine. The team has made the source code openly available, accelerating reproducibility and clinical deployment.