Grammar-Based Lexicon Extension for Aligning German Radiology Text and Images

Proceedings of Recent Advances in Natural Language Processing, pages 105–112, Hissar, Bulgaria, 7-13 September 2013. Grammar-Based Lexicon Extension for Aligning German Radiology Text and Images Claudia Bretschneider1,2 1Center for Information and Language Processing, University Munich claudia.bretschneider.ext @siemens.com Sonja Zillner2 2Corporate Research, Siemens AG sonja.zillner @siemens.com Matthias Hammon3 3Department of Radiology, University Hospital Erlangen matthias.hammon @uk-erlangen.de Abstract For efficient diagnosis processes, the mul- titude of heterogeneous medical data re- quires seamless integration. In order to automatically align radiology reports and images based on the pathological anatom- ical entities they describe, a preceding sen- tence classification is necessary. However, the lexical resource used has to contain se- mantic information about the pathological classification of each entity. We introduce an approach to extend medical lexical re- sources with pathology classification in- formation and, at the same time, with new classified vocabulary. Our algorithm is based on a semi-supervised learning algo- rithm and incorporates a semantic context- free grammar combined with a RadLex- based lexicon. 1 Introduction In radiology, the health status of a patient is de- scribed using a multitude of formats. During the examination process, a radiologist creates ma- chine readable descriptions such as radiology im- ages, dictated reports about the image findings and written texts. Although, most of the radiology data are related via the anatomical entities shown or de- scribed, there is no link between them, since the information pieces are stored in distributed sys- tems. This absence of links between the items is hindering the radiologist’s workflow. Especially when reading reports, radiologists want to refer- ence back from the described finding (in the text) to the correlating body location (in the images). Without automatically created links, this resolu- tion is obviously time-consuming when dealing with images taken with modalities that deliver a mass of stacked images. Today, radiologists add alignment information to the text that names the image that contains the described findings. But still, the resolution of these textual links requires manual interventions to find the correct image and detect the described finding in the image. To simplify this workflow, we introduce a mechanism that automatically aligns pathological anatomical entities in radiology text and images based on semantic annotations. Figure 1 shows our concept of linking anatomical concepts from image and text: Both the images and the texts are annotated with the anatomical concepts that they describe. Combining annotations with the same RadLex ID (RID), the link from one format to the other can be established. As a result, the radi- ologist can easily navigate from the pathological Leber [liver] (RID58) described in the text to the correlating position in the images. For the integration, the necessary semantic an- notations of the images have been made avail- able as a result of a previous project (Seifert et. al., 2009; Seifert, 2010). In order to align these RadLex-based annotations with anatomical enti- ties described in radiology reports, our text anal- ysis system has to annotate the texts with RadLex- based annotations, too. Our established mecha- nism operates in two …

Content truncated to 500 words