RSNA 2016

Abstract Archives of the RSNA, 2016


SSC08-08

Development of a Novel Bayesian Network Interface for Radiology Diagnosis Support and Education

Monday, Nov. 28 11:40AM - 11:50AM Room: S402AB



Po-Hao Chen, MD, MBA, Philadelphia, PA (Presenter) Nothing to Disclose
Suyash Mohan, MD, Philadelphia, PA (Abstract Co-Author) Grant, NovoCure Ltd; Grant, Galileo CDS, Inc
Tessa S. Cook, MD, PhD, Philadelphia, PA (Abstract Co-Author) Nothing to Disclose
Ilya M. Nasrallah, MD, PhD, Philadelphia, PA (Abstract Co-Author) Nothing to Disclose
R. Nick Bryan, MD, PhD, Philadelphia, PA (Abstract Co-Author) Stockholder, Galileo CDS, Inc; Officer, Galileo CDS, Inc
Emmanuel J. Botzolakis, MD,PhD, Philadelphia, PA (Abstract Co-Author) Nothing to Disclose
CONCLUSION

A prototype web-based interface (ARIES) was developed that streamlines interaction of radiologists with BNs. With further development and validation, we anticipate this could provide Radiology diagnosis and educational support.

Background

Bayesian networks (BNs) are forms of artificial intelligence that have shown promise for Radiology diagnosis support. Taking as input imaging and clinical key features (KFs) extracted by radiologists, BNs can output probability-ranked differential diagnoses (DDx) and suggest further imaging or testing to constrain the DDx. Moreover, because BNs illustrate probabilistic relationships between KFs and DDx, they offer a unique approach to Radiology education that emphasizes “bottom-up” diagnostic reasoning (i.e., DDx given KFs), as opposed to more traditional “top-down” approaches (i.e., KFs given DDx).

Evaluation

To translate BNs into clinical and educational practice, we developed ARIES (Adaptive Radiology Interpretation and Education System), an open-source, web-based interface that allows Radiologists to interact with expert-developed BNs representing various imaging domains (e.g., Neuroradiology). ARIES utilizes a commercially available BN backend (Netica, Vancouver, Canada) wrapped in a Java server, and was created using JavaScript, JQuery, and HighCharts. ARIES was developed in close collaboration with practicing radiologists, intended for use alongside a traditional PACS workstation.

Discussion

In Clinical Mode, ARIES displays buttons corresponding to relevant KFs. As KFs are selected, two sets of probability-ranked DDx are continuously updated ("radiographic DDx," based on imaging KFs alone, and "clinical DDx," using both disease prevalence and clinical KFs). Embedded sensitivity analysis highlights the next most discriminating KFs after each selection. In Education Mode, trainees are prompted to review clinically proven cases from an internal teaching file. After entering KFs and providing a DDx, automated feedback is provided comparing agreement between trainee- and expert-extracted KFs, and between trainee- and BN-generated DDx. ARIES also offers machine learning functionality, updating BN probability tables in real-time as cases are submitted to the interface.