The basic goal of this feasibility study was to verify the applicability and suitability of the RadLex ontology in the reporting of gliomas, which has not been performed so far. Only under this basic condition may the benefits of a broad application of this standardized terminology in the given context be realized, namely the reduction of variation and increase of clarity in radiology reports . By abstracting the inter-individual diverse vocabulary to a conceptual level in terms of specific RadLex items, comparability of findings (e.g. follow-up versus initial MRI exams) in clinical day-to-day practice as well as large-scale data aggregation for epidemiological research purposes and health care quality management measures would be facilitated because the heterogeneity and thus complexity of linguistic processing would be markedly reduced. In concrete terms, consistent application of this terminological standard in glioma reporting would make it possible to extract the language-encoded information by natural language processing software and to network it with other electronic health systems within the framework of an integrated infrastructure. This would be beneficial to brain tumor patients due to improved and less ambiguous interdisciplinary communication and additionally contribute to the development of comprehensive databases, which could be used for epidemiological research, health care decision making, and training applications. One example of such successful use for educational purposes is the integration of RadLex in medical content-based image retrieval algorithms to help radiologists make the right decisions through comparison with similar cases from an adequately annotated database . As for the gains related to research, RadLex coding of the specific terms allows radiological findings to be easily translated into other languages unequivocally and without loss of content, which may facilitate transnational research collaborations and help patients in times of increased mobility. In addition, more comprehensive search results can be obtained through targeted database queries when terminology is standardized. Taken together, all these advances are only possible if the underlying terminology is applicable to the respective radiological context. It was the aim of our exploratory study to evaluate this basic requirement for the reporting of gliomas. We therefore transposed 20 consecutive free-text brain MRI reports on patients with newly diagnosed GBM into the current version of the RadLex ontology. More than 95% of descriptive terms used in these reports were satisfactorily covered by means of verbatim, synonymous, or combinations of existing RadLex items. More than a decade ago, Marwede and colleagues investigated a preliminary RadLex version with regard to indexing of thoracic computed tomography reports and observed a degree of completeness of 84% for this radiological subspecialty . Furthermore, an analysis of a large set of published radiology reporting templates demonstrated a partial or complete match between 2.509 extracted unique terms and corresponding RadLex elements in a proportion of 67% . A comprehensive evaluation of more than 385,000 radiology-centric figure captions gathered from 613 peer-reviewed medical journals revealed the best term mapping performance for RadLex compared to five other biomedical ontologies despite comprising fewer items . Since then, the Radiological Society of North America has made huge efforts to elaborate and extend this controlled vocabulary. Therefore, the current fourth version of RadLex includes more than 46,600 distinct entities . Nonetheless, there remains an inherent risk of fragmentariness of the RadLex ontology owing to its top-down construction process by expert committees with experience in various radiological subdomains . Fortunately, several attempts have been undertaken in the past to overcome this structural downside by means of automatic software based extraction of terms from different large-scale sources including a corpus of Pubmed repository articles as well as an enormous set of 270,540 free-text mammography reports [10, 23]. The latter study was performed with the aid of natural language processing, which could be an option to expand other radiology domains as well by using this complementary approach. Shore and colleagues scrutinized books, radiological articles, dictionaries, and biomedical webpages for names and synonyms of imaging signs that were subsequently integrated into RadLex’s “imaging observation” section to improve its applicability . Following this bottom-up concept, we manually analyzed consecutive free-text brain MRI reports of a cohort of high-grade glioma patients. The assessment displayed a small fraction of terms (< 5.0%) that could not be attributed to a specific RadLex entity or RID combination, respectively. First of all, the lexicon lacks options to communicate the exact magnitude of a specific tumorous lesion. Apart from coarse size descriptors such as small, medium, and large or alternatively less than 10 mm, 10–19 mm, and 20 mm or greater, there is no further possibility of refining this important characteristic . Especially in view of patients suffering from GBM, spatial tumor dimensions are a well-established independent prognostic factor with respect to overall survival together with other clinicopathological features such as greater extent of resection, younger patient age, better physical condition, and eventually O-6-methylguanine-DNA-methyltransferase promoter methylation status [25, 26]. Bearing in mind this clinical implication, we would welcome the implementation of more detailed size descriptors in the upcoming updates on the RadLex vocabulary. Furthermore, the concept of ‘blood–brain barrier disruption’ was stated in one of the surveyed radiological documents and could not be matched adequately by any RID. On the one hand, it is not possible to visualize the blood–brain interface directly via conventional MRI—what you might see is a T1 enhancement due to extravasation of contrast medium into the extracellular space as a result of a disruption of this neuroprotective barrier, but not the barrier itself [27, 28] and therefore it could be argued that this concept may be expendable. On the other hand, a primary range of the RadLex application involves indexing of large databases of radiological free-text reports for educational and research purposes, which necessarily requires the highest achievable degree of completeness [3, 6]. Hence the issue of implementation of a specific term into RadLex should be decided upon preferably by the fact how deeply ingrained this concept is in our radiological everyday communication. A substantial part of items that were represented in the analyzed GBM MRI reports and not attributable to any specific RadLex entities could be delineated as pictorial signs, such as finger-shaped brain edema or garland-like tumor enhancement. Despite a certain lack of objectivity of such descriptions, these kinds of figurative terminological elements have been appreciated and widely accepted by radiologists for interpretations in neuroimaging right from the beginning of the era of clinical computed tomography application [29, 30]. As early as in the mid-seventies of the twentieth century a pioneer report on the diagnostic possibilities of cranial computed tomography made references to tumor-related white matter brain edema “producing finger-like shapes” in a large cohort of patients . If the RadLex terminology is meant to be a common lingua franca for the radiological community, it should be oriented towards the principles of general language evolution. The Duden dictionary, first published by Konrad Duden in 1880, provides the preeminent language resource of the German language and states the authoritative rules regarding utilization of German language. It is regularly updated and the editorial decision on inclusion of a particular word or phrase is mainly based on its frequency and longevity of use . Because of the widespread adoption and long-term usage of the above mentioned figurative radiological terms we propose the augmentation of the RadLex vocabulary with these elements. This approach would be well in line with the general policy pursued by the Duden curators in terms of everyday language. The synergy of a continuously expanded and updated RadLex terminology adapted to everyday practice and a set of essential morphologically describable features, as developed in the VASARI project for brain tumors by The Cancer Imaging Archive (TCIA), has the potential to sustainably improve the quality, precision, and communication of MRI reporting of GBM .
This study is not without limitations. The monocentric study design as well as a relatively small sample size, which was chosen due to the tremendous efforts required for manual data extraction, make up downsides of the survey potentially compromising its generalizability. On the other hand, the clearly defined eligibility criteria and thorough scrutiny of all consecutive MRI reports by two independent neuroradiological raters assure an explicit statement on the issue of the applicability of RadLex in GBM MRI reporting. Moreover, the included MRI reports were authored by a large group of ten experienced neuroradiologists all contributing their specific reporting style and vocabulary, which may increase the variability of the terms used and thus tests the basic practicability of the RadLex ontology studied in this specific neuro-oncological context.