Computer Vision News - April 2020

2 Summary Research 6 Where is the unitary normalized gradient field of a sphere, and the parameters p and a respectively affect the interaction between the points and the shape of the field (centripetal/centrifuge). The map is concatenated to the initial feature maps of the encoding part of the 2 nd block but it is also included on the final segmentation layer, due to the skip connections in the U-Net shape. The weight map hence influences all the weights of the model and its output. The map also reappears in the second term of the loss function defined as: Training ℳ ℳ ℒ attraction = 1 − ∑ ((ℳ > ) ∘ ) ∑(ℳ > ) This looks at common segmentation points between the predicted mask and the weight one within a region of interest controlled by γ . The total loss then combines the attraction termwith a term based on the Intersection over Union (IoU) between the ground truth mask and the predicted one. ℒ = 1 ℒ IoU + (1 − 1 )ℒ The images used for training are LDCT (low-dose CT) scans with annotations made by four different radiologists. The information provided with the dataset includes an inter-observer agreement level which accounts for the segmentation variability, and the nodule texture. The training is then made through pairing the same nodule with different segmentations and according to the details grouped in the table below. The training was performed in 2 steps: - training of the 1 st block using only until validation loss has reached a plateau ℒ IoU - training of both blocks using the total loss until no change is recorded for 5 epochs (no change for 3 epochs) ℒ IoU