Unsupervised Feature Loss

UFLoss, or Unsupervised Feature Loss, is a patch-based unsupervised learned feature loss for deep learning (DL) based reconstructions. The UFLoss provides instance-level discrimination by mapping similar instances to similar low-dimensional feature vectors using a pre-trained mapping network (UFLoss Network). The rationale of using features from large-patches (typically 40×40 pixels for a 300×300 pixels image) is that we want the UFLoss to capture mid-level structural and semantic features instead of using small patches (typically around 10×10 pixels), which only contain local edge information. On the other hand, the authors avoid using global features due to the fact that the training set (typically around 5000 slices) is usually not large enough to capture common and general features at a large-image scale.