Computer Vision News - December 2018

central part of the chamber is described by another mathematical formula, with different parameters. After all these components are described mathematically, they can be blended using mathematical functions to create a shape. By comparing these shapes to database from CTs, we show that these mathematical methods can describe almost any left atrium shape from the database of real patient data. Being this a deformable model, it can be manipulated using a small number of parameters instead of many thousands of points or mesh vertices. We only need some tens of numbers to describe a wide variety of shapes. This model can be used to generate many artificial atria that look plausible, each with its own score: this can be used to augment the dataset of real patient atria to train a neural network. The model is based on a denoising autoencoder attempting to reconstruct the input given to it. The architecture is fully connected, but when training a denoising autoencoder one should set some percentage of the input to zero. The result obtained is a noisy input. Then, instead of just using noise, we deleted some area of the data, either by using an intersection of the sphere of the atria or by simulating the path of a catheter introduced to perform an ablation, as would be done by a physician during a real procedure. This part takes only one minute or so and is followed by the attempt to reconstruct the shape. Alon Baram of Tel Aviv University says : “ What is challenging here is that the network needs to imagine how the 3D shape will look like using the statistics that it has learned during training. For this, we introduced a new regularization that helps the network learn smooth shapes. This is done by using a spatial gradient of the weights, as low as possible. This has not been tried on any network: unlike a convolutional neural network where weights are shared along the space, this network is fully connected but constrained to learn only smooth shapes. ” We can say that this collaboration combines two technologies in a creative manner: one is a parametric model, describing the shape using a limited number of parameters to get a geometrical and mathematical representation of it; the second is a neural network. The first technology is used to generate artificial models to provide data augmentation needed for training the neural networks to do the 3D reconstruction task. Project 13 Computer Vision News A project by RSIP Vision Left: rigid phantoms; center: CAD to path registration; right: phantom registration - red is recorded path and blue is template Take us along for your next Deep Learning project! AlonBaram Tel AvivUniversity MosheSafran RSIPVision