Computer Vision News - September 2020

Research 8 Nevertheless, GANs have also some drawbacks. Those include the sensitivity in the hyperparameter tuning, but also a recently researched topic: “ mode collapse ”. This is the failure of the GAN discriminator to get confused as the generator over- optimises for a particular discriminator . The result is that the generators only go through a subset of output types. In this work, intensity inhomogeneities in MR imaging are exploited to create realistic adversarial examples for the specific problem. The intensity is not uniform and this is described by the following: Φ is the bias field and I the intensity of the image. The bias field is made of low frequencies and varies across the image. The repeated random sampling of this bias field would be computationally inefficient; therefore, the approach was to target the weakness of the network using a min-max game described by: To keep this short, the point here is to construct an adversarial bias field by maximising the distance between the original prediction and the prediction after perturbation (measured by Dcomp). The parameter θ of the network is then optimised with the aim of minimisation of the distance between the initial prediction and the prediction after the generated adversarial bias attack (fseg) . The final task is to optimise the segmentation network by computing the Dcomp and applying in to regularise the network. MR images of the left ventricular myocardium from the ACDC public dataset were used to evaluate the results of the network against state-of-the-art methods using 100 subjects and dividing to 5 groups including a healthy group and four with cardiac abnormalities . Cropping of 128x128 was applied as a preprocessing step and all the images were bias-corrected using the N4 algorithm. The N4 bias field correction algorithm corrects low-frequency intensity non-uniformity present in MR imagingdata known as a bias or gain field. Random affine transformation, as well as elastic, horizontal and vertical flipping, were applied to the images. More training details can be found in detail on the paper, as well as the segmentation used, basically a 2D U-Net using 2D image slices as its input.