Martin Huber PhD, Presenter: Employee, Siemens AG

Sascha Seifert PhD, Abstract Co-Author: Employee, Siemens AG

Alexander Cavallaro MD, PhD, Abstract Co-Author: Nothing to Disclose

Gerhard Kohl, Presenter: Employee, Siemens AG

The advent of initiatives like DICOM structured reporting and RadLex are manifestation of the need to more formally capture the content of radiology reports and therefore medical images. Complementing these forthcoming standards with an image parsing system will result in improved reporting workflows.

In today’s PACS/RIS systems, radiological images and reports are linked on a study level only. In order to link anatomical structures in CT volumes directly to the corresponding findings in the radiology report, both images and reports are parsed automatically and annotated using common vocabulary from clinical ontologies.

An image parser scans the 3D CT volumes for the main organs (lung, heart, liver, kidneys, bladder, prostate, spleen) and 32 salient anatomical landmarks (e.g. bronchial bifurcation, bottom tip of sternum). The position of organs and landmarks is saved in a database. A text parser scans the reports for names of anatomical structures making use of the RadLex thesaurus and the foundational model of anatomy (FMA). FMA represents the human anatomy as a hierarchical structure of 70000 distinct entities, linked by 170 relations like e.g. “part-of”. We removed 47000 entities not relevant for radiology and added the 32 anatomical landmarks. If the anatomical term found in the text is known to the image parser, the corresponding image location is looked up in the database. If not, anatomical landmarks are searched in the hierarchy below and their geometrical centre is used as image location. In case of missing landmarks, hypernyms are searched until one is found whose location is known. Vice versa, text passages in the radiology report are highlighted whenever the focus of the 3D image viewer is within a known anatomical structure.

Automated image parsing efficiently extracts image content and makes it actionable for intelligent image search, fast navigation and alignment with text documents. Specifically the linking to radiology reports promises workflow improvements. Increasing the number of anatomical structures detected in the image volumes is necessary to improve the accuracy of the determined positions.

Huber, M, Seifert, S, Cavallaro, A, Kohl, G, Automated Alignment of Anatomical Structures in CT Volumes with Radiology Reports.  Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9011703.html