Computer Vision News - January 2021

5 eliminates the border pixel discontinuity , enhances the training/validation/ testing accuracy by achieving less computational parameters. In the case of axial T2W images though image-to-image approach isn’t used directly because of the heterogeneous prostate capsule boundary, pixel intensity, prostate location and other parameters. To minimise those drawbacks, a radiomics-based deeply supervised 2D U-Net network was used. The radiomics approach is used to describe a collective data extraction from images to analyse those in a quantitative fashion. Features that can represent the prostate region are size, shape, intensity, volume etc. The aim of the pipeline presented in Fig 1. is to extract a large number of such features using the axial T2W images. The flowchart of the proposed pipeline is shown in Fig 1b. and it shows the data input, pre-processing stages as well as the radiomics-based deeply supervised network stage. The radiomics feature maps are 1024 and made of flattened filters of the T2W MR 30 x 30 channel feature maps. There are 1024 feature maps in the encoding part whereas the number of feature channels are 512 in the decoding part which are concatenated with decoding part feature maps in order to get the segmented image details. Figure 1: The proposed radiomics-based 2D U-Net pipeline, describing with colored arrows the respective layer operations and input and output sizes. Prostate Lesion Segmentation