Computer Vision News - January 2021

7 For the training, a stochastic gradient descent algorithm is used with the Adam optimiser. The learning rate was set after experiments to 2x10-4 and is the speed the input radiomics features map is learned. A couple of selected cases from the paper are shown in Fig. 3. The proposed pipeline shows a lot of improvement, even in a visual inspection in comparison to the DS-CNN segmentation. Prostate Lesion Segmentation Figure 3: Prostate lesion segmentation from two selected cases from the paper (from left to right): axial T2W images, the ground truth segmentation, DS-CNN segmentation, and proposed pipeline. Figure 4: The proposed network achieves a mean score of 0.9176 and it tops the current state-of-the-art.