Computer Vision News - August 2020

3D-RADNet 31 Best of MIDL 2020 On the top row, after applying a sliding window method for patch extraction, the selection of the region of interest containing the liver is completed by 3D-RADNet. Once the regions are correctly selected, they are used as input for the second part of the pipeline (2nd row), where the decoder for the segmentation network is based on the VNet structure and optimized with Dice loss. For this task, the LiTS challenge dataset available here was used, including 131 CT scans, previously interpolated into 5mm spacing, and split into 70% for training, 15% for both validation and testing set. The training was carried out several times by freezing different residual blocks of the network and with random subsets of the training data. These attempts are shown below, and they demonstrate that the best results were obtained from the network trained with all encoding 3D-RADNet layers frozen (DICE = 90.0%), while the lowest segmentation performance was obtained with the network trained from scratch and initialized with 3D-RADNet weights without freezing any layers.