Self super resolution for hepatic vessel CT segmentation
Vincent Jaouen, Zhihan Wang, Pierre-Henri Conze, Dimitris Visvikis · 2023
Accurate hepatic vessel segmentation from abdominal CT images is a major requirement for hepatic diagnosis and surgery. However, it is a very complex task due to the intricate nature of the vascular anatomy and the limitations of CT imaging. A growing number of work advocate for the critical importance of pre-processing in the success of deep supervised image segmentation techniques. For instance, abdominal CT images are often reconstructed into anisotropic voxelized volumes with a typical ratio of 2 to 5 between in-plane and through-plane voxel sampling, which hampers analysis and vessel detection in the low resolution plane. In this work, we study whether self super resolution (i.e. learning-based super resolution relying on the anisotropic volume only) could be used to improve liver vessel segmentation performance in abdominal CT images. This way, we avoid resorting to curated super resolution training sets that are difficult / impossible to collect in practice. We leverage a self super resolution approach previously proposed for brain MR images, SMORE, and demonstrate its interest for fine vessel reconstruction. In particular, we show that reporting segmentation results in terms of global overlap metrics such as Dice cannot faithfully account for segmentation performance improvements with this type of data. For this reason, we also consider the recently proposed clDice metric accounting for the connectedness of the segmentation output. Using a nn-UNet segmentation benchmark, we show that we significantly improve on average segmentation results using the proposed self super resolution stage, both qualitatively and quantitatively.