E2698. Man Vs. Machine: A Single Center Experience of Artificial Intelligence Detection of Pulmonary Emboli
Authors
Alexander Song;
Staten Island University Hospital
Deven Kulkarni;
Staten Island University Hospital
Varun Mehta;
Staten Island University Hospital
Background
Artificial intelligence (AI) has been claimed as a revolutionary tool and as a potential threat for the radiologist. The rise of image recognition software outside of medicine has garnered great interest in the application of AI within radiology. Although applications in image acquisition and radiologist workflow have been studied, perhaps of greatest interest is image interpretation by AI. We present a single center experience with AI detection software implementation and detection of pulmonary emboli on contrast enhanced chest computed tomography (CT).
Educational Goals / Teaching Points
Briefly review AI and applications within radiology. Discuss AI implementation and utilization by radiologists at our institution. Discuss AI accuracy and pitfalls in detection of pulmonary emboli at our institution. Explore future directions and potential improvements for AI pulmonary emboli detection software.
Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
AI detection of pulmonary emboli at our institution has shown to be overly sensitive and thus, susceptible to many artifactual causes of apparent filling defects within pulmonary arteries on contrast-enhanced chest CT images. Radiologists showed greater accuracy than AI in detection of true pulmonary emboli and were better in accounting for technical factors such as contrast bolus timing, breathing motion, and streak artifact. A survey of our radiologists regarding the usefulness of the AI detection software yielded varying responses.
Conclusion
Although it is a useful tool to alert the radiologist to the possibility of a pulmonary embolus, AI was shown to be inferior to radiologists in detection of true pulmonary emboli at our institution. Further improvements in AI software are needed to account for varying technical factors between different chest CT examinations. These limiting factors are likely common to AI interpretation of other anatomic imaging studies and there is much room for improvement in diagnostic accuracy.
