Computer Vision News - April 2020

3 Summary iW-Net 5 Hence, the network proposed by the authors poses a new possible key to solve this problem. IW-Net performs both automatic and interactive segmentation of lung, it has been evaluated on the public LIDC-IDRI dataset (available here ) and its source code can be found here . It could be a fundamental tool for increasing early diagnosis of lung cancer. iW-Net iW-Net is a less dense version of the 3D U-Net. A first auto-encoder outputs an automatic segmentation (orange layer in figure below, 1 st block) which then undergoes a second auto-encoder that produces the final corrected segmentation (blue layer in figure, 2 nd block). The novel element lies in the pixel-wise weight map (green layer) which is used both as a feature map of the network and as a loss function term. is defined as the absolute value of a vector field W that moves between two user introduced points, Q0 and Q1, and its magnitude is higher in the region between the centres of these points and lower elsewhere. This is used as a metric for the network to understand which region is of more interest for the segmentation. The points are defined as below: = (−1) ∇ |∇S| ℳ ℳ