Biomedical Ontologies to Guide AI Development in Radiology

Area of Science:

• Radiology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Deep learning has spurred significant interest in AI for radiology, including image analysis, report manipulation, and intervention planning.
• Successful AI applications require integration with medical knowledge, which is currently limited in scale for widespread use.
• Ontologies offer a structured, computer-readable format for representing domain-specific knowledge.

Purpose of the Study:

• To explain how ontologies can support AI research and development in radiology.
• To highlight the role of ontologies in guiding emerging AI applications within the field.
• To demonstrate the potential of ontologies in enhancing AI's impact on patient care and research.

Main Methods:

• Defining ontologies as computer-based knowledge representation systems that capture term semantics through relationships.
• Discussing the characteristics of biomedical ontologies, including causal, part-whole, and anatomic relationships.
• Referencing existing ontologies like RadLex as examples of applied knowledge representation in radiology.

Main Results:

• Ontologies provide a human-readable and machine-computable method for encoding medical knowledge.
• They enable AI systems to leverage medical knowledge at a larger scale, improving development and application.
• Ontologies can guide diverse AI applications such as natural language processing, image analysis, radiomics, and treatment planning.

Conclusions:

• Ontologies are essential for the successful and responsible development of AI in radiology.
• They facilitate the integration of medical knowledge into AI systems, enhancing their utility and reliability.
• Leveraging ontologies will accelerate AI's contribution to biomedical research and clinical practice in radiology.