Search

Deep Learning approaches for automatic segmentation of organs from CT scans and MRI are providing promising results, leading towards a revolution in the radiologists' workflow. Precise delineations of abdominal organs boundaries reveal fundamental for a variety of purposes: surgical planning, volumetric estimation (e.g. Total Kidney Volume – TKV – assessment in Autosomal Dominant Polycystic Kidney Disease – ADPKD), diagnosis and monitoring of pathologies. Fundamental imaging techniques exploited for these tasks are Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), which enable clinicians to perform 3D analyses of all Regions of Interests (ROIs). In the realm of existing methods for segmentation and classification of these zones, Convolutional Neural Networks (CNNs) are emerging as the reference approach. In the last five years an enormous research effort has been done about the possibility of applying CNNs in Medical Imaging, resulting in more than 8000 documents on Scopus and more than 80000 results on Google Scholar. The high accuracy provided by those systems cannot be denied as motivation of all obtained results, though there are still problems to be addressed with. In this survey, major article databases, as Scopus, for instance, were systematically investigated for different kinds of Deep Learning approaches in segmentation of abdominal organs with a particular focus on liver, kidney and spleen. In this work, approaches are accurately classified, both by relevance of each organ (for instance, segmentation of liver has specific properties, if compared to other organs) and by type of computational approach, as well as the architecture of the employed network. For this purpose, a case study of segmentation for each of these organs is presented. [ABSTRACT FROM AUTHOR]