Computer Vision News - April 2021

9 SAX images only, from the 1st sub-cohort of the TOF dataset, including 193 sets (144 used for training – 39 for testing) and the ACDC data (75-25). The transformation module was trained in a four-fold cross-validation using only the 2nd cohort of the TOF dataset - 81 patients with pairs of AX and SAX images, shuffled and split into 60-21 for training and validation. To summarize, the authors use the method described above to lead four main experiments, whose performance is measured using the Dice score, Volume difference and Hausdorff distance. These are: 1. Domain gapwhen SAXmodel is applied on axial slices without domain adaptation 2. Upper limit with respect to a direct regression on the transformation parameters 3. Baseline domain adaptation approach 4. Complete domain adaptation approach – where Dice and HD are considerably improved with respect to all three. According to the DICE score, results show that there is an increase in performance of 8%(LV) - 2%(MYO) - 25%(RV) on the 4 th experiment compared to the 1 st ; of 2-4% on all classes from 2 nd to 4 th and of 24%(LV) - 18%(MYO) - 22%(RV) from 3 rd to 4 th . What does Tal Arbel wish Computer Vision News for our 5th anniversary? Find out on page 28! Unsupervised Domain Adaptation from ...