Computer Vision News - April 2020

3 Summary iW-Net 7 The authors carried out two main experiments: 1) analysis of the output of the 1 st block of the network and how it relates to the expert’s agreement level; 2) evaluation of the impact of artificially generated user inputs. For the second experiment, they compared the output of the 1 st auto-encoder with the full iW-Net output (Corrected (Cr) segmentation). As metrics, they used an Average IoU: Experiments and performance Optimizer Adam Learning rate 0.001 Batch size 8 Augmentation Yes (random rotations, translations, flips and zooms) Split of training and validation 80%-20% Validation 5-fold cross validation Hyperparameters Random search (100 steps) ̅̅̅̅̅̅̅̅̅ = 1 ∑max ( ( , , ̂ , ), ( , , ̂ )) =1 where N represents the number of radiologists annotating the nodule and j is the nodule considered each time, and the 3D IoU is defined as: ( , ̂ ) = ( ⋂ ̂ ) ( ⋃ ̂ )