E4912. External Validation Study on CTA-Based Detection of Aortic Dissection with an Artificial Intelligence Algorithm
Authors
Roshan Fahimi;
Massachusetts General Hospital
Parisa Kaviani;
Massachusetts General Hospital
Emiliano Garza Frias;
Massachusetts General Hospital
Lina Karout;
Massachusetts General Hospital
Keith. J Dreyer;
Massachusetts General Hospital
Subba Digumarthy;
Massachusetts General Hospital
Mannudeep K Kalra;
Massachusetts General Hospital
Objective:
Diagnosis of suspected and unsuspected aortic dissection is important to decrease the risk of aortic rupture, which carries high mortality and morbidity. Our study represents external, standalone validation of an artificial intelligence (AI) algorithm for detection of aortic dissection on dynamic contrast-enhanced chest or chest-abdomen CT examinations.
Materials and Methods:
This retrospective study included 267 patients (mean age 64 ± 14 years, 109 women and 158 men) from two quaternary care hospitals who underwent dynamic contrast-enhanced chest or chest-abdomen CT examinations. All examinations were performed on either single source (64/512-row GE) or dual-source (192/384-row Siemens) scanners. Patients were identified with the keyword of aortic dissection on a commercial radiology reports search engine (Nuance mPower). We recorded information related to presence and location of dissection. Thin-section DICOM images (1–1.25 mm) were deidentified, exported, and processed with AI algorithm (Avicenna.AI, U.S. FDA, and EC approved) on an offline analytic platform (CARPL) to obtain information on the presence and location of aortic dissection. Data were analyzed with ROC using AUC for diagnosis of dissection.
Results:
AUC, sensitivity, and specificity for the diagnosis of aortic dissection were 0.87 (95% CI 0.93–0.82), 0.83, and 0.92, respectively. There was no difference in AI performance across the two institutions (site A: AUC 0.88 [95% CI 0.90–0.95], sensitivity 0.81, specificity 0.95 vs. site B: 0.87 [95% CI 0.80–0.95], 0.85, 0.90) (p > 0.5). AUC, sensitivity, and specificity for men and women were similar as well (men: 0.85 [95% CI 0.78–0.92], 0.83, 0.88; women: 0.91 [95% CI 0.82–0.98]). AUC, sensitivity, and specificity across the two CT vendors were 0.86 (95% CI 0.78–0.94), 0.75, 0.97 versus 0.88 (95% CI 0.81–0.95), 0.92, 0.85, respectively.
Conclusion:
The assessed AI algorithm was effective and generalizable for detecting aortic dissection on chest and chest-abdomen CTA. Rapid triage and diagnosis of aortic dissection with AI can be helpful in expeditious reporting of CT with critical findings.
