The present effort focuses on developing a method for assisting the representation of radiological reports, written in simplified natural language, in a standardized and content recovery prone structure, such as DICOM Structured Report. Sample reports were collected and have being analyzed. The work is currently in process, but an intermediary representation was already reached and is being evaluated by physicians to attest the accuracy of the results.
Most health care institutions have a precious legacy of clinical reports written in natural language, or simplified grammatical structure. Unfortunately content based retrieval of information from these reports is inefficient due to the peculiarities of natural language. This prevents institutions from sharing clinical records without waste of precious time and resources.
The present research effort focuses on the development of methods based on
knowledge about normal findings in radiological examinations [
In the same way there must be a more appropriated technique to solve each task, where several tools of AI and other areas of knowledge are combined in only one intelligent system that manages the efforts to solve tasks.
Clinical report, especially in radiology contains information concerning a patient’s medical condition. However, a great percentage of this information is not structured, for it’s free text based, that consequently makes it more difficult to search, analyze, summarize and present.
Previous studies have shown the potential benefits of medical structured data for
the practice, research and medical teaching. The information’s structure can be used
to help organize and improve the medical record presentation [
Other systems that seek to enlarge the semantic contents of ontology to guide the
knowledge discovery on database and also analyze the variables types that the user
checks [
One fact that deserves to be singled out is the lack of works directed to
applications in the Portuguese Language, face to the fact that most terminologies are
found mostly in English, having a few versions to other languages as German, French
and Spanish. Portuguese is the eighth most spoken language on the planet, third
among occidental languages, after English and Spanish [
SNOMED is a widely accepted terminology and infrastructure designed to enable the sharing of health care knowledge, across clinical specialties and sites of care. It contains, preferred medical terms and concepts, consisting of more than 144,000 terms and term codes divided into eleven: Topography; Morphology; Function; Living Organisms; Chemicals, Drugs, and Biological Products; Physical Agents, Activities and Forces; Social Context; Diseases/Diagnoses; Procedures; and general Linkage Modifiers. SNOMED is available in several languages including German. Unfortunately there is not yet an available translation of the SNOMED terms for Portuguese. Thus, in order to make this experiment possible, the basic terms for thoracic radiological exams were translated by collaborator Brazilian physicians.
The DICOM SR standard sets out rules that define how structured documents that
contain health information should be composed, stored and transmitted. These make
use of a controlled terminology; which enhances the results of content based retrieval
[
The sample thoracic reports, 315 of them in German and 7719 in Portuguese, were analyzed. In the sample reports a common structure was identified. A heading containing the identification information such as of the physician, patient, and institution in which the report was emitted, followed by a description of the techniques used to perform the exam, such as type of procedure. The main part of the report is the body in which the conclusions drawn from the exam can be found. This will be the focus of the effort to interpret, retrieve and codify content.
The language used by physicians to describe findings in radiological examinations
in both the institutions can not be considered indeed natural language. It is rather a set
of short sentences normally without the use of verb, escaping the linguistic
conventional rules. These, often either affirm of deny the existence of a malformation
or alteration in an anatomical structure, the presence of strange bodies or illnesses,
adding observations regarding the localization, shape or appearance of the anatomical
structure. For each type of clinical report of radiological examination there are
specific anatomical structures of interest [
The subject of most sentences found in thorax radiological reports are anatomical structures such as mediastinum, lung, trachea and others. The subject is often followed by adjectives stating the position, or the part of the organ that was observed, for instance surface or right. Adverbs, describing morphological information follow, such as normal or reduced. To this, might follow information regarding diseases, like for example, the existence of emphysema and adverbs indicating the degree of severity of the disease. For the analysis and structuring process ware used natural language processing techniques, lexical, syntactic, semantics and of speech analysis combined with information from the specialist physicians. Because the medical language and usually language have a great difference, normality the language in the clinical reports are affirmative or negative phrases with few or without verb. 5
The sample radiological reports from Dataflex and word documents base were
converted to simple text ASCII standard representation in order to disregard
formatting information. Afterwards, the GNU Aspell free tool [
For the natural language processing, the text file is separated in sentences observing punctuation and end of line commands. Later the words are separated for the lexical analysis. Using a former developed dictionary, the words are classified by their grammatical class, such as substantive, adjective and others. The sentences and words are the input for the syntactic and semantics analysis. Each word has a function in the sentence, but we search first the terms for anatomical structures, diseases and morphology are then matched to the appropriate SNOMED codes.
After the first tests of classification, there was found a great amount of adjectives that derived from substantives and for that reason had a huge importance, needing then a special treatment, for were directly referred to anatomies. See the table 1. For each sentence without an anatomical SNOMED matching term, a speech analysis is used, the algorithm searches an anatomical term from the previous phrases. When no term is found, the code of anatomical term from the clinical report’s title is used. Earlier the use of specific rules for each type of exam was used, but due to the complexity and difficulty to make a great number of rules, computationally were studied the resemblance and differences among the medical reports and the sentence structures contained on those. Despite the diversity of sentences as well as in numbers of words as in the way of position or construction, the most important information to be extracted are the anatomical region (where?), the diseases or alterations (what?), and the qualifications of those (how?). Additionally can be used the position into anatomical region and general modifiers, such as right, left, bilateral, surface, superficial and others.
The terms for anatomic regions are searched in the “Topography” branch of SNOMED. If a word regarding position is along with the anatomic region, that also can be converted to one of the terms in the “General Linkage and Modifiers” branch of the SNOMED hierarchy, and the diseases or alterations (what?) can be converted in categories “Morphology“, “Diseases/Diagnoses“ or “Function“. In this part of the process we obtained the support from the Teachers and Resident Doctors in General Surgery from HRAV (Hospital Regional Alto Vale).
In association with the overall are found the terms of qualification (how?) such as: normal, intact, anatomic, etc, and can also be combined with the denial as: without, absent, etc, that indicate the absence of problems, or their intensity such as: minimum, light, acute, chronic, etc. In the following example of analysis, on the first original sentence of the chart, there are three anatomic regions and one qualification, that would generate three sentences according to the where and how items, in a relation N anatomies to 1 qualifier, and the word “bronchi” also applies to “lobar”.
On the second sentence is noticed the relation of 1 to 1 and on the last sentence is found a relation of 1 anatomy to N qualifiers, besides finding the pleural derived object that needs to be converted for the substantive Pleura added by the adjective surface, that can be found respectively in the “Topography” and “General Linkage Modifiers” branches of SNOMED. For each anatomical SNOMED term and adjective in a phrase, a new sentence must be created; this process can be visualized on the Table 2. The language used by physicians to describe findings in radiological examinations in both collaborator hospitals is a simplified restricted one. Nevertheless, there is much research still to be done before a reliable representation of such reports in a standard such as DICOM SR can be obtained.
The objective of this work is not to produce software capable of performing a complete automatic translation of clinical reports written in natural language to a standardized form, but to develop an approach to facilitate de process of structuring documents, so that these are better suited for content based retrieval.
Although it is currently being tested using reports written in Portuguese and German, the same method slightly adapted is expected to function with other languages. The results so far reached are at the present time subject to the evaluation of the physicians to attest the validity and suggest better approaches. The authors thank the physicians of the Chirurgic Residence of the Upper Itajaí Valley Regional Hospital in Brazil for the translation of Radiological SNOMED terms to Portuguese and for the anonymized sample reports provided for the analysis. The Buddenbrock Blasinger und Benz Radiological Clinic, in Mainz, Germany, contributed providing anonymized sample reports. We also thank the German Academic Exchange Service -DAAD- for the scholarship number A/03/42304 granted to one of the authors. online