Computer Vision News - April 2021

10 Research One might wonder why the volume difference appears among the metrics. It is extremely important here as the authors take care to mention that in AX images, the average RV volume appears 51 ± 15 ml larger than in the corresponding average SAX volume. A significant difference between the two is found performing the T-test between RV ground truth volumes (AX-SAX). This variation between the shape and volume of RV in the two axes might be more than what would be expected to be found due to the different grid orientations. And in conclusion, this means that the network might systematically predict smaller RV regions than the expert contour in axial view, because it is trained only on native SAX ground truth contours. As promised, this paper delivers a full pipeline to segment axial images just leveraging the much bigger dataset of short-axis scans and labels. It is really worth looking at one example where the complete model has been applied to a held- out AX image stack. For this image, the obtained Dice scores for each volume are 0.9274 (LV), 0.8707 (RV) and 0.6605 (MYO).