Surface classification on a 3D cardiac ventricular model using machine learning

This work improves techniques for the classification of cardiac ventricular surfaces on a polygonal surface mesh in the context of small datasets. This task is reduced to a multi-class classification of a point on a surface mesh. Machine learning models are trained to classify the polygonal mesh vertex based on the values of a signed distance function in the neighborhood of this vertex. Machine learning models are compared, including FCNN, U-Net, and ResNet neural nets, and classifiers from the scikit-learn library. In addition to accuracy measures, the suitability of the classification results for constructing the biventricular coordinate system is assessed. A graph algorithm is proposed for correcting potential classification errors and its effectiveness is demonstrated. Models using neural networks are found to be the most effective. Less resource-demanding models that exhibited comparable performance are the Random Forest and Support Vector Classifier with Stochastic Gradient Descent.