Entwicklung qualitativer Bilddeskriptoren zur Klassifikation morphologischer intrakranieller CT- und MRT-Befunde

BACKGROUND: Medical knowlegde is subject to a rapid evolution, making its organisation in transdiciplinary knowlegde systems difficult. This is contrasted by the postulation that clinicians should ground their decisions on scientific evidence that, in turn, requires optimized access to information. In an ideal case, knowlegde based information systems can be confronted with findings from individual patients and support differential diagnostic or therapeutic decisions. AIM: In this work, a step towards such a system in the field of neuroradiology, we suggest the fragmentation of textbook knowledge into prototypical descriptive findings, also referred to as descriptors, that in combination with exemplary images can function as independent cells of knowledge. MATERIAL AND METHODS: In the department of neuroradiology at the Max-Planck-Institute of Psychiatry, Munich, we retrospectively reviewed 141 cranial MRT and CT cases with confirmed diagnoses from the categories of neoplasms (N=42), vascular diseases (N=49), inflammation and demyelination (N=13), neurodegeneration (N=8) and other (N=4). Both by image and image report review features typical and relevant for the respective diagnosis were exctracted. This selection was complemented by descriptors from over 280 image descriptions and text passages from neuroradiological textbooks and review articles. Descriptores were translated to english, condensed and arranged in a hierarchical scheme that is referred to as neuroradiological description scheme (NRDS). After recursive refinement and comparison with neuroradiological experts, the scheme was integrated into the multidisciplinary knowledge base N-Expert (Institute of Applied Sciences in Medicine, Munich) and used to construct entries in a hyperlinked electronic textbook, structured reports for individual cases and structured descriptions of single example images. The NRDS was investigated for its coverage of established clinical and radiological thesauruses. Furthermore, we explored the type and depth of the text-image-linkage of established online neuroradiological systems, and searched the literature for alternative knowledge representation systems that could serve as basis for structured image description and neuroradiological diagnosis decision support systems. RESULTS AND DISCUSSION: The NRDS with about 150 hierarchically organized descriptors can be used to implement core findings from cranial MRT or CT into a digital textbook. It can further be used as navigation instrument to guide the user efficiently to textbook passages, exemplary clinical cases or exemplary images and by this, support medical education. Decision support is promoted by accelerated access to a neuroradiological phenomenon and by the possibility to search the database along different search axes with each different granularities. The parallel access via an anatomical classification, a diagnosis access and via neuroradiological and clinical characteristics is innovative and prevents inefficient browsing. The linkage between prototypical, single findings and image examples in the N-Expert framework allows also for a fast recruitment of material for teaching purposes. A deeper, expert-like diagnostic logic, however, was difficult to express by the scheme. Specifically, its strictly monohierarchical structure were identified as drawback. In addition, an additional entity between elemental findings and diagnoses is essential to aggregate elemental findings to syndrome-like intermediate classes. In addition, explicitly defined relationships between the descriptors in the NRDS would be of advantage. The description of elemental findings using the NRDS was possible in principle. While its expressiveness can be expanded by including other source domains, the neuroradiological case descriptions did not yet represent a well readable structured report. In part, this was caused by the overcomplex and uncomfortable representation of anatomical information. Literature review, however, showed, that up to now no system has been presented that is sufficienly expressive for clinical reporting and that can in parallel be used as interface to clinical information systems. A limitation of this work is the lack of a quantitative, formal evaluaton of coverage, semantic precision and diagnostic relevance of the scheme. OUTLOOK: The NRDS, however, can serve as systematic starting vocabulary that can be further adapted to alternative knowledge representations as for example semantic networks. In addition, epidemiological data could be integrated as well as recently described clustering and visualisation methods to reduce the multidimensional feature space of complex findings and to identifiy prototypes and similarity between cases. By such transformations of the scheme, its classification and decision support potential can can be further exploited.