A scoping review of the knowledge graph used in medical imaging analysis

A recent scoping review provides readers with aggregated documentation of the state-of-the-art knowledge graph applications in medical imaging analysis to encourage future research.

The review was published in Health Data Science, a Science Partner Journal.

Deep learning methods have great potential in analyzing medical images with algorithms trained to identify abnormal regions and tissue variations. Knowledge graphs can provide structured domain information to help computers build more intelligent systems and solve real-world problems. 

In this work, Yifan Peng, assistant professor with the Department of Population Health Sciences, Weill Cornell Medicine, and colleagues conducted an article review to identify the increasing trend of incorporating knowledge graphs in medical imaging analysis. 

“We hope our article can provide readers with aggregated documentation of the state-of-the-art knowledge graph applications in medical imaging to encourage future research.” says Peng as he shares the intent of this work.

The literature has extensively explored knowledge graph applications in biomedical informatics, but this review focused on medical imaging analysis, a specific and unique perspective. 

To fill the vacuum in literature, the authors conducted a systematic search in five databases for the relevant articles, screened and reviewed the selected articles, and analyzed knowledge graph applications in the medical imaging domain. 

During this process, 21 eligible articles were included in the systematic analysis. This relatively small number of relevant articles may suggest the unmet need for cross-disciplinary collaborations.

Medical imaging analysis tasks described in these articles center around disease classification, disease localization, report generation, and image retrieval. One of the most important messages from this project is that incorporating prior knowledge into medical imaging analysis tasks has proved its effectiveness, shares Song Wang, coauthor and student of The University of Texas at Austin.

In addition, their analysis reveals limitations among the existing works, namely a limited amount of available annotated data for some supervised tasks and weak generalizability to other tasks. 

The authors then identified potential future directions, such as exploring semi-supervised frameworks, different fusion methods, and task-agnostic models for better generalizability; and introducing more sophisticated graph structures to model disease relationships with increased sophistication. 

Indeed, the proven effectiveness of knowledge graphs in medical imaging analysis tasks encourages researchers to consider utilizing prior knowledge in related projects.