Computer Vision News - November 2020

3 Ilya Kovler at MLMIR 35 There are previous works in this field, but they are mostly based on statistical shape models or based on deep neural networks with standard input representation and loss functions . RSIP Vision’s solution is based on a deep neural network. It looks like a standard neural network architecture, but the novel part is the way we represent the input X-rays. We introduce a dimensional enlargement approach that combines two bi-planar x-rays back-projections of each pair of corresponding epipolar lines into a single two-channeled epipolar plane. It also creates two additional loss functions , which make the result much more accurate than in other methods. “For a supervised loss function, we added to the voxelwise cross-entropy loss a spatial weight to give more importance to the challenging bone edge regions,” Ilya explains. “For each training sample, we define a spatial 3D Distance Weight Map.” The other loss is an unsupervised reconstruction loss employed to align the neural network prediction bones probability map with the input X-ray images. This loss encourages the network to use the available information from the input images that can actually be used in inference time, where no supervision is available. Best of MICCAI 2020