Automatic Re-orientation of 3D Echocardiographic Images in Virtual Reality Using Deep Learning

In 3D echocardiography (3D echo), the image orientation varies depending on the position and direction of the transducer during examination. As a result, when reviewing images the user must initially identify anatomical landmarks to understand image orientation - a potentially challenging and time-consuming task. We automated this initial step by training a deep residual neural network (ResNet) to predict the rotation required to re-orient an image to the standard apical four-chamber view). Three data pre-processing strategies were explored: 2D, 2.5D and 3D. Three different loss function strategies were investigated: classification of discrete integer angles, regression with mean absolute angle error loss, and regression with geodesic loss. We then integrated the model into a virtual reality application and aligned the re-oriented 3D echo images with a standard anatomical heart model. The deep learning strategy with the highest accuracy - 2.5D classification of discrete integer angles - achieved a mean absolute angle error on the test set of 9.0 degrees. This work demonstrates the potential of artificial intelligence to support visualisation and interaction in virtual reality.