Disease Ontology
The Disease Ontology (DO) is an extensive, hierarchically organized vocabulary of 12,694 human diseases that provides a framework to identify relationships between diseases and phenotypes, genotypes, and various other disease attributes. DO provides semantically consistent annotations that allow one to compare diagnostic evaluations, treatments, and patient-care data over time and between studies. DO incorporates concepts and extensive cross-mapping from standard clinical and medical terminologies, such as Medical Subject Headings (MeSH), International Classification of Diseases (ICD), Online Mendelian Inheritance in Man (OMIM), and the National Cancer Institute Thesaurus. The ontology provides a resource to connect genetic and phenotypic information related to human disease.
Human Phenotype Ontology
The Human Phenotype Ontology (HPO) describes phenotypic features of hereditary, congenital, and acquired diseases using a structured and controlled set of terms. Although focused initially on monogenic diseases—about 50,000 annotations connect HPO terms to 4779 diseases in the OMIM database of genetic disorders—HPO now includes features of more than 3400 common non-Mendelian disorders. HPO terms can have more than one parent in the phenotypic hierarchy: for example, podagra (gout of the big toe) has parent terms gout and abnormality of the foot. HPO has been linked to OMIM and to the Orphanet Rare Disease Ontology (ORDO) to increase the interoperability of phenotypic knowledge in rare diseases. One can match clinical information to phenotypes at varying levels of specificity in the ontology's hierarchy to formulate differential diagnoses; for example, a clinical record describing a "short 2nd toe" would be linked to the HPO term short 2nd toe and its more general terms, short toe, short digit, and abnormal digit morphology.
Integration of Ontologies
Individual ontologies can serve as building blocks of broader, more general knowledge resources. By integrating related ontologies, knowledge can be shared and reused across domains. The US National Library of Medicine's Unified Medical Language System (UMLS) Metathesaurus seeks to provide semantic integration of concepts across ontologies and vocabularies; a single concept unique identifier (CUI) in the UMLS Metathesaurus may refer to concepts in several component vocabularies. As described above, the Foundational Model of Anatomy (FMA) forms much of the basis of the anatomical terms in RadLex. Knowledge of radiological differential diagnosis in RGO has been integrated with SNOMED CT, RadLex, DO, HPO, and the ORDO. This integration allows one to pose new, abstract questions that relate diseases and their imaging phenotypes such as, "Which gastrointestinal system diseases may cause an abnormality of the genitourinary system?" (Fig. 4).
Limitations
Although ontologies provide a powerful framework to organize information, they have several limitations both in their design and application. With more than 800 ontologies available through NCBO BioPortal (https://bioportal.bioontology.org), it can be challenging to discover new or relevant ontologies. Mappings between these ontologies are sparse and few are included in the UMLS Metathesaurus. SNOMED CT, LOINC, and ICD-10-CM are part of the Metathesaurus, but, notably, RadLex is not. Licensing of ontologies can pose another barrier: the US federal government licenses SNOMED CT for use in the USA, but the ontology is not openly available worldwide. Uneven or incomplete development of ontologies can limit their use: many ontologies, including RadLex, are built either through volunteer effort or are part of a larger initiative. Only a minority of RadLex terms include a definition; new concepts that enter the vocabulary of radiology may not yet be incorporated, and there are ongoing questions of the scope and purpose of any ontology. As ontologies grow larger, it is increasingly challenging to assure that they are maintained appropriately and that the ontology's axioms are logically consistent.
Few ontologies other than SNOMED CT and LOINC have been incorporated broadly into commercial products, and solutions adopted in industry often resort to simpler, non-standards-based textual information. Towbin et al. have highlighted the widespread and substantial limitations of traditional free-text body part labeling in DICOM and the importance of an ontology for body part labeling to enable enterprise imaging. Smith et al. review many of the challenges of ontologies for biomedical imaging, and propose a coordinated suite of ontologies to address the needs of biomedical imaging. There may be lack of incentive in the commercial space as organizing or labeling data ontologically requires more work and maintenance; it is unclear what might motivate such adoption.
Applications
Natural Language Processing
Natural language processing (NLP) has been applied widely in radiology, and can perform more efficiently than human experts with near-human accuracy. Many clinical and research applications rely on automatic recognition of medical concepts in unstructured text, and the accuracy of recognition can affect the analysis of electronic health records. The mining of medical concepts is complicated by the broad use of synonyms and nonstandard terms in medical documents. Ontologies have been shown to improve performance of NLP applications.
Ontologies can support radiology NLP applications by providing a comprehensive vocabulary of a domain, along with the terms' synonyms, abbreviations, and relationships. Ontology's synonyms, abbreviations, and related terms can help identify diagnoses and other entities in radiology reports. For example, RGO incorporates 2957 synonyms (e.g., hepatoma and HCC for hepatocellular carcinoma) that can improve recall of an NLP search routine and can also help coalesce the findings into a single unique entity. Links between ontologies can offer additional synonyms; for example, the RGO term LEOPARD syndrome has DO synonyms: Capute-Rimoin-Konigsmark-Esterly-Richardson syndrome, generalized lentiginosis, Gorlin syndrome II, lentiginosis profusa syndrome, Moynahan syndrome, multiple lentigines syndrome, Noonan syndrome with multiple lentigines, and progressive cardiomyopathic lentiginosis. Hierarchically organized information allows one to search intelligently for classes of findings; for example, a search for phakomatosis encompasses more specific terms such as neurofibromatosis, NF1, and tuberous sclerosis.
Text de-identification tools often use algorithms that attempt to redact people's names by identifying words that are capitalized. Radiology reports, however, commonly contain eponyms for devices, findings, and diseases that include capitalized proper nouns (e.g., Foley catheter, Kerley B lines, Wilms tumor). Thus, de-identification tools may require domain-specific vocabulary so that these terms are not inappropriately redacted from the record or replaced by arbitrary names as in de-identification techniques such as "hiding in plain sight". More than 800 of these imaging findings (such as Rigler sign) have been incorporated into RadLex. Connections between ontologies can identify additional similarly capitalized proper noun synonyms as alluded to in more detail above. Although modern neural network–based approaches have reduced the need for feature engineering, exposure to controlled vocabularies can still supplement these techniques and limit the need for extremely large sets of training data.
In addition to search-and-replace tasks and named-entity recognition, ontologies can supplement deep neural network–based techniques by creating word embeddings based on ontology terms. An embedding technique converts words into vectors, or numerical representations, based on neighboring and co-occurring words in a text corpus. SNOMED2Vec, a word embedding model that uses SNOMED CT to augment machine-defined relationships, can outperform models that are not based on ontologies because it strengthens complex medical relationships such as "breast + adenocarcinoma → invasive ductal carcinoma".
The relationships in an ontology enable more robust medical reasoning. The multiple hierarchical levels of terms in an ontology enable connections that go beyond word-list associations (e.g., for a synonym or proper noun detection) to make an end model more medically relevant. Researchers have explored semiautomated approaches to learn ontologies from text. NLP performance was improved on a gene/protein synonym detection task by adding formal ontological knowledge without modifying the word embeddings; the ontology provided context that effectively induced term variability while reducing ambiguity. Biomedical ontologies evolve continually to become more robust and to accommodate new imaging techniques and findings. Ontologies can augment NLP, and NLP can be used to analyze free-text corpora to find gaps and to suggest terms to add to existing ontologies.
Image–Based Machine Learning
Image–based machine learning can be applied to a virtually infinite space of potential applications across diseases, findings, and imaging modalities. Early successful image–based machine learning applications have focused narrowly on particular findings or diseases; a key challenge is to identify the applications that will have the greatest impact on radiologists' performance and patient care. Ontologies can help guide investigators towards fruitful avenues for research. For example, musculoskeletal radiology alone entails 635 differential-diagnosis listings in RGO, such as abnormal odontoid process and premature craniosynostosis, with their interlinked imaging findings and diseases. The differential-diagnosis listing for abnormal odontoid process links to 37 possible causes, including ankylosing spondylitis and systemic lupus erythematosus. Similarly, in thoracic imaging, RGO includes 286 imaging findings caused by 2958 diseases (Table 3). Understanding these linkages can guide researchers to focus on findings that are intertwined with diseases to increase the chance of substantive clinical impact.
These hierarchical relationships and mappings can then help translate findings that are detected using image-based techniques to diseases or other entities which may strengthen an image–based machine learning model and help with one key area of machine learning, explainable or interpretable AI. Image-based models may use saliency maps ("heat maps") or activation maps to explain which image features contributed most strongly to certain findings in an exam. Ontologies and their relationships can then help synthesize these findings to propose causal relationships or subsequent diseases and outcomes and can show the mappings that led to these conclusions. For example, an image-based AI model may detect cirrhosis; ontological knowledge can suggest what might have caused the finding (alcohol abuse, viral hepatitis, etc.) and possible consequences (portal hypertension, hepatocellular carcinoma, etc.).
One of the most challenging and time-consuming parts of developing image–based machine learning tools is labeling data. Early approaches typically required cumbersome manual labeling by domain experts, but automated or semiautomated techniques that generate large corpora of slightly imperfect training data can approach and nearly match the performance from carefully manually curated labels, although weakly labeled datasets generally require significantly more labels to approach the same level of performance, Ontologies can support the development of datasets to train or test image-based AI models by identifying mentions of conditions or imaging findings in radiology reports.
High-quality NLP models can facilitate important image–based machine learning applications and their evaluation. Marrying an image-based model that detects certain findings or entities with an NLP model that can accurately identify those findings in the report could be used to automate peer review at a scale far greater than what is feasible with current manual methods. Machine learning tools used by a practice should be continually monitored for performance and for safety events; parsing report data for the findings suggested by the image-based model can help quantify the performance and potential errors of both AI systems and humans.
Radiomics
Radiomics can link the characteristics of radiological images with a patient's genotype, gene expression profile, and prognosis. The use of ontologies can help identify the scope of diseases or imaging findings related to a particular radiology exam and guide the integration of radiomics features with specific diseases. The Image Biomarker Standardization Initiative (IBSI) has defined uniform, reproducible approaches to compute 174 quantitative features from medical images. The IBSI radiomics features have been integrated with segmentation algorithms and imaging filters into the Radiomics Ontology, which is published through NCBO BioPortal. An ontology-guided radiomics analysis workflow has been developed to capture features from imaging data to facilitate research and clinical translation of radiomics. As radiomics methods become more commonplace and sophisticated, one can imagine that particular computational features will become associated with various conditions, and such information could be incorporated into a diagnostic ontology.
Radiation Therapy Planning
Ontologies can aid in planning, which is the task of identifying optimal sequences of actions to achieve a specified goal. An example of planning is found in mapping applications that can identify the fastest route to a given destination. In diagnostic radiology, a planning task might be to identify the most cost-effective imaging procedures to answer a clinical question. A common planning task in radiation oncology is to select optimal radiation ports to treat target lesions while sparing nontarget organs.
Two ontologies have been developed in radiation oncology. The Radiation Oncology Ontology (ROO) represents clinical data in the radiation oncology domain; it incorporates 1183 classes and 211 properties between classes. The authors used the ontology to query information from various data structures without a priori knowledge of the schemas of the underlying relational databases. The Radiation Oncology Structures (ROS) ontology comprises 417 anatomic and treatment planning classes that describe commonly contoured structures for radiation treatment planning.