Meeting Banner
Abstract #1586

Database-Guided Detection and Segmentation of Organs in MR FastView Localizers for Automatic Scan Planning

Neil Birkbeck1, Nathan Lay1, Jingdan Zhang1, Artem Gritsenko1, Jens Guehring2, S. Kevin Zhou1

1Imaging & Computer Vision, Siemens Corp., Corporate Technology, Princeton, NJ, United States; 2Imaging & Therapy Division, Siemens Healthcare, Erlangen, Germany

We investigate automatic multi-organ localization in large FoV localizer datasets acquired using a fast continuously moving table technique (syngo TimCT FastView, Siemens AG, Erlangen, Germany). We developed a fast learning-based detection and segmentation method for 6 organs including liver, heart, lungs, and kidneys. The automatically identified anatomical information allows for precise automated scan planning based on an organ or structure of interest. We compare the accuracy of our detection and segmentation routines on ground truth annotations from 196 full body MR scout scans and achieve segmentation accuracy that are within the voxel spacing on average and are computed in under 8s.

Keywords

accuracy achieve achieved adding allow allows among anatomical annotations appearance attention automated automatic automatically axial body boundary cardiac chamber classifier classifiers coefficients collection come complete computational computed computer continuously corp corporate correspondence cost cross database datasets dependencies describe detection developed deviation displaced distance distribution division done dual easily efficiency efficient estimation evaluated extra fast features final fine fold framework full function gained gives ground guided heart hierarchical identified illustrates impression isotropic kidney kidneys larger learning leveraging limited liver localization localizer localizers locally location lung lungs make manner marginal mesh model modeling models moving network object operator organ organs orientation overall overlap overview percentile pipeline planning position post precise predicted processing qualitative quality regularized remove reproducibility resolution robustness routines scale segmentation segmented semantic sequential shape signed slices space spatial specify statistical steerable step structure subsequent surface surfaces system table takes technology therapy trained truth unknown utilize validation variation velocity vision volume volumes