RadiO: A Prototype Application Ontology for Radiology Reporting Tasks

Abstract

We introduce RadiO, a prototype application ontology for the support of electronic radiology reporting. This application ontology is implemented in Protégé and comprises three layers: 1. a radiology report layer, capturing observations made on patient examinations through the use of a controlled vocabulary of the radiographic imaging domain (RadLex), 2. an imaging domain ontology, representing knowledge about image entities and their image features, and 3. a reference ontology for anatomy (Foundational Model of Anatomy), representing canonical anatomical knowledge. The aim of this prototype is to support the identification of image features of image entities and their use in diagnostic interpretations, as well as to provide a basis for structured reporting applications in the domain of medical imaging.

Introduction

The construction of biomedical ontologies has focused on the representation of knowledge in specific biomedical subdomains like anatomy¹, surgery procedures², genetics³, or pathology⁴. However, to date, no ontology concerned with the diagnosis of diseases has been developed. In the domain of medical imaging we propose that there are two reasons for this: firstly, until recently, no widely accepted standardized terminology has been available, and secondly, no ontological knowledge model of disease diagnosis has been constructed because of the particularly intractable problems posed by the diagnostic act for ontological analysis.

We developed an ontological framework for radiology reporting which interfaces a recently published controlled vocabulary used for reporting on radiographic images (RadLex⁵), via an intermediary ontology of the image, to the Foundational Model of Anatomy (FMA)¹, a reference ontology for anatomy. The aim of this application ontology is to build a knowledge base of imaging ‘findings’ and their interpretation as diagnoses in the domain of medical imaging.

We call this framework, and the application ontology which is supported by it, RadiO. Using RadLex term categories, RadiO’s reporting layer provides terms for the reporting of observations based on images of bodily entities. The lexical Elements in the resulting electronic radiology report are subsequently structured according to an image ontology that we have designed. Through this intermediary image ontology, with its set of basic ontological relations, we established an interface between the FMA and radiological reporting tasks in clinical practice.

Diagnostic Domains

In the medical domain many different tests exist which are deployed to prove a diagnosis. For each test the diagnostic value can be measured statistically. In the domain of medical imaging the diagnostic value of an imaging examination depends on the type of pathology and its features accessible by imaging examinations. Some pathological conditions, e.g. pneumothorax, can be easily observed on images whereas others, e.g. bronchial carcinoma can be difficult to diagnose. Diagnostic difficulty in medical imaging is produced by the fact that some pathologies insufficiently exhibit image features and others do not present a set of unique image features⁶. In the domain of clinical radiology, image observations are usually referred to as ‘findings’ which vary from simple image features (e.g. opacity) to complex diagnosis (e.g. broncho-alveolar carcinoma).

Clinical radiology

Clinical radiology is a rapidly evolving discipline dedicated to producing images of the human body through imaging techniques such as computed tomography (CT), magnetic resonance imaging (MRI), or ultrasound (US). Radiology departments nowadays are highly computerized environments concerned with digital acquisition and processing of images, as well as the electronic storage of images and reports by which radiologists document image observations. Yet, despite the increasing reliance on computers in radiology departments today, the content of radiology reports, which serve as the basis for the communication of diagnostic results, has still not been significantly affected by information management technologies such as ontologies⁷.

Some lexicons have been developed for specific subdomains of clinical radiology like the Fleischner Glossary for thoracic imaging⁸ or the BIRADS classification system for mammography reporting⁹. However, the Radiology Society of North America (RSNA) is now developing a comprehensive lexicon of radiological terms with the aim to unify terms used in radiology reporting and to serve as a platform for interfacing these terms to other biomedical data sources. RadLex provides the basic means for developing tools capable of analyzing large amount of report data and realizing the scientific advantages that may come with such analyses, such as determining the large-scale reporting and interpretation patterns of radiologists.

Application and reference ontologies

A reference ontology is a reusable and generalizable resource that is designed to meet the needs of any knowledge based application requiring structured information¹. Application ontologies, in contrast, are constructed for particular user groups and serve specific practical tasks related to the relevant domain.

In general, application ontologies, accessing knowledge in a reference ontology, need only employ those entity types and relations necessary for executing specific tasks. However, to date, there are few examples where the distinction between reference and application ontologies has been adopted⁹.

Aiming towards providing interoperability with already developed ontologies in the medical domain, our application ontology complies with the OBO Relation Ontology¹⁰ and accesses knowledge in existing reference ontologies (FMA).

Material and Methods

RadLex: A standardized radiological lexicon

The lexicon consists of 7466 terms classified in 9 major term categories¹⁰. The majority of RadLex terms are subsumed under the categories of Anatomic Location (3290), Finding (1925), and Modifier (762). The term category of Finding defines pathological conditions or diseases frequently diagnosed as a result of imaging observation. The term category of Modifier contains different types of features or classifiers used to characterize findings.

Domain of the image

In clinical practice, the main task of a radiologist is to interpret images of the body for diagnostic purposes. For any given image, a radiologist has to determine if this represents a normal or pathological condition on the side of the patient and document the associated signs accordingly. If pathological, the radiologist must report what discernable image features provide evidence for this diagnosis. Thus, the assertions made by radiologists in their reports primarily refer to the image itself, and only secondarily to the organs these images are images of.

For example, the assertion “normal liver” in a radiological report indicates that the appearance of this particular liver in this particular image is normal. This is important to note since not all pathologies are susceptible to imaging, such as cancer on the cellular level, and a liver that appears normal in an image may in fact not truly be so. Thus, for any application ontology dealing with the application of imaging techniques to the human body, it is important that bodily entities and their appearances are kept in separate (but interrelated) ontological domains: that of the human body and, in projected form, on the image.

Aligning RadLex term categories with external ontologies such as the FMA requires creating an intermediate ontology of the image entities that RadLex terms refer to. In the domain of medical imaging we distinguish three image entity types: anatomical image entity, pathological image entity and image feature. Anatomical image entities link to anatomical entities like those defined in canonical anatomical knowledge bases like the FMA. Pathological image entities link to diseases and diagnoses like those defined in disease classification systems like ICD 10¹¹ or SNOMED⁴ whereas image features are attributes of anatomical or pathological entities exhibited on an image (see Figure 1).

Figure 1.

Aligning RadLex term categories, the imaging domain, and the body domain.

Results

Application ontology for radiology reporting

The application ontology is implemented in Protégé and currently consists of three layers: 1. a radiology report model, 2. an imaging domain ontology, and 3. an anatomical knowledge base (FMA).

Radiology report model

The radiology report model consists of a radiology report class in which observations on images can be instantiated. According to the DICOM standard each reporting class contains information about the patient (e.g., name, ID, clinical history), the imaging modality employed (e.g., CT, MRI, US), and the results of the examination. Based on the clinical situation at hand (imaging modality, body region examined, and the clinical question), appropriate report items are pre-defined for the report class. For example, a CT-examination of the lung in a patient with suspected pulmonary embolism has the anatomical image entity pulmonary arteries specified as a template item which must be evaluated by the examining radiologist.

Imaging Domain Ontology

The entities of the imaging domain relate to each other by two basic relations:

• c has_location c¹: basic location relation holding between image entity instances. [Where c is an image feature or pathological image entity and c¹ is an anatomical image entity at which it is situated.] E.g.: This reticulonodular pattern has_location lung.

• c has_feature c¹: an upper-level property relation holding between two image entity instances. [Where c is an anatomical or pathological image entity, c¹ is an image feature attribute].

In the imaging domain ontology three different types of image features are distinguished: general features describe attributes like extent (e.g. partial, complete), trend (e.g. increasing, decreasing), or timing (e.g., acute, chronic). Visual features are features describing the size (e.g., enlarged, small), composition (e.g., hemorrhagic, calcified), or morphology (e.g., atrophic, reticular). Modality-related features are signal features specific for an imaging modality (e.g. for CT hyperdense, isodense, hypodense).

Anatomical image entities

The class anatomical image entity is linked to the class anatomical entity of the FMA through the image_of relation. By instantiating a subclass of anatomical image entity, e.g. pulmonary arteries, the relation image_of is automatically instantiated, which links to anatomical entities from the FMA, thus importing anatomical knowledge into RadiO. Elements of the class image feature are linked to entities of the class anatomical image entity through the relation has_feature.

Pathological image entities

Entities of the class pathological image entity are linked to the elements of the class image feature by the has_feature relation and to the elements of the class anatomical image entity by the has_location relation.

Image features

In the imaging domain ontology the relation has_feature is used to link elements of the class image feature to specific pathological and anatomical image entities. This relation has several subtype relations for different types of image features. Image features are located at the appropriate anatomical image entity through the relation has_location. Figure 2 displays the relations between different image entity types and the FMA.

Figure 2.

Relations between image entity types and the FMA

Reporting and interpreting image entities

For the reporting of imaging observations, image entity types are instantiated in the report class. We differentiate 4 reporting patterns frequently found in radiology reports. Reporting of normal anatomical conditions is done by evaluating the condition as normal. Image observations might be followed by an interpretation:

1. Evaluation of anatomical image entities, e.g.: lung parenchyma has_evaluation normal.

2. Anatomical image entities and their image features, e.g.: thyroid gland has_size enlarged.

3. Image features of anatomical image entities and their interpretation as pathological image entities, e.g.: lymph node has_composition calcified, has_interpretation silicosis.

4. Pathological image entities and their image features and location, e.g.: pulmonary thromboembolism has_location lumen of right pulmonary artery, has_extent complete.

For each image entity instantiated in the reporting class, the reporting radiologist has to specify the anatomic location. For example, if a radiologist wishes to describe an observation in the upper right lobe of the lung the relation has_location needs to be instantiated in the reporting class. The description of image features is performed by instantiating one or more of the subtype relations of has_feature, like has_shape, has_margin, or has_composition. This set of relations may be followed by an interpretation using the relation has_interpretation. Figure 3 displays an observation instance of a report class describing a set of different features observed in the liver followed by an interpretation.

Figure 3.

Reported Observation composed of three different Image Feature Instances (Visual Feature, Modality-related Feature, General Feature) and their corresponding feature templates with reported attributes (Margin, Shape, Density, Amount) followed by an interpretation.

Discussion

Biomedical ontologies are increasingly emerging with the aim of structuring knowledge and allowing electronic processing of biomedical data. Even if modeling of report content in the radiology domain has been performed¹², no ontology concerned with the diagnosis of disease has been developed. We present an application ontology constructed for the electronic annotation of image features to image entities observed in imaging examinations. The aim of this application ontology is to construct a knowledge base of image features at specified anatomic locations and their interpretation as diagnoses.

In our implementation we differentiate a radiology reporting layer, an imaging domain ontology, and an anatomical knowledge base. To take account of the image entities produced by different imaging techniques and the diagnoses which can be inferred from those images, we distinguish explicitly between the body and the image, two separate ontological domains which are aligned via the image_of relation.

In accordance with RadLex we distinguish three upper-level entity types in our imaging domain ontology: anatomical image entity (linking to the FMA), pathological image entity (linking to disease classification systems) and image feature (characterizing image entites). We have defined a set of basic relations to annotate image features of image entities and implemented a set of commonly used reporting patterns. By doing so, reporting patterns and specific image observations can be tracked to discover which diagnoses are inferred from imaging observations by the use of which sets of entities.

Conclusions

We have developed RadiO, a prototype application ontology for radiology reporting tasks, which aligns a controlled imaging vocabulary (RadLex) to a reference ontology for anatomy (FMA). The aim of this implementation is to support the structured reporting of image observations and to build a knowledge base concerning how image entities are used in the process of diagnosing diseases. Future work will include the empirical validation of the ontology and the composition of more detailed reporting templates to facilitate daily reporting tasks in radiology, as well as the alignment of imaging diagnoses to emerging disease ontologies.