Computer Vision News - August 2020

Research 30 Best of MIDL 2020 3D-RADNet The 3D-RADNet is a classification network, adapted from the ResNet50 architecture, that takes 3D inputs of size 48x192x192 and outputs a single answer for each of five classes, including the aforementioned image modality/sequences, view, contrast, plus scan coverage and slice spacing. The main parameters are recorded in the table. Note that the first of these matches the difference in classes. Parameter Type Class Activation function I. Softmax modality/sequence, view and contrast II. Sigmoid scan coverage layer III. Linear slice spacing Loss function Cross-entropy loss and Root-mean squared - Optimiser ADAM optimiser - Performance On the classification task , 3D-RADNet scored 99.4% in determining the view and 84.8% in the contrast agent classification, while the others all achieved more than 90% in the testing set. An MSE of 3.1mm was obtained for slice spacing regression. For the segmentation task , carried out to demonstrate transfer learning performance, the pipeline is displayed in the figure below.