Daniel Raskin, Gal Yaniv , Chen Hoffmann, Eli Konen
Department of Diagnostic Imaging, Sheba Medical Center, Tel Hashomer, Affiliated to Sackler School of Medicine, Tel Aviv University, Israel
2Department of Neurosurgery, Mount Sinai Hospital, USA
PURPOSE: To evaluate the specificity and sensitivity of Aidoc’s deep learning technology in flagging pathological hyperdense intracranial lesions in non-contrast head CT (NCHCT).
METHOD AND MATERIALS: A dedicated computer aided (CAD) deep learning algorithm was designed for the detection of pathological intracranial hyperdense lesions (PIHL) on a NCHCT. This study is a retrospective review of consecutive NCHCT examinations in an emergency department of a single center during a week in April 2018. All examinations were reviewed and tagged for PIHL by a resident and a senior neuroradiologist. The results were matched with the outcome of Aidoc`s flagging. The sensitivity and specificity of CAD was compared with the gold standard – a senior neuroradiologist examination report.
RESULTS: Total of 160 cases were reviewed during a single week period. According to the ground truth, a total of 34 positive scans (21.2%) and 126 were negative (78.8%) were included. Three out of the 34 positive scans were not detected by the Aidoc solution, resulting in an overall sensitivity of 91.1% (CI: 0.76-0.98%, P<0.05). Out of 126 negative scans, 3 was flagged as a positive, resulting in an overall specificity of 97.6% (CI: 0.93-0.99%, P<0.05). Positive predictive value was 91.1 % (CI: 0.76-0.98%, P<0.05), while negative predictive value was calculated as 97.6 % (CI: 0.93-0.99%, P<0.05). Accuracy was 96.2% (CI: 0.92-0.98%, P<0.05).
CONCLUSION: Aidoc’s deep learning technology demonstrated high accuracy in flagging PIHL. Integration of CAD for detection of hyperdense intracranial finding presents promising specificity and sensitivity.