Analysis of head CT scans flagged by deep learning software for acute intracranial hemorrhage

Purpose To analyze the implementation of deep learning software for the detection and worklist prioritization of acute intracranial hemorrhage on non-contrast head CT (NCCT) in various clinical settings at an academic medical center. Methods Urgent NCCT scans were reviewed by the Aidoc (Tel Aviv, Israel) neural network software. All cases flagged by the software as positive for acute intracranial hemorrhage on the neuroradiology worklist were prospectively included in this assessment. The scans were classified regarding presence and type of hemorrhage, whether these were initial or follow-up scans, and patient visit location, including trauma/emergency, inpatient, and outpatient departments. Results During the 2 months of enrollment, 373 NCCT scans were flagged by the Aidoc software for possible intracranial hemorrhage out of 2011 scans analyzed (18.5%). Among the flagged cases, 275 (72.4%) were positive; 290 (77.7%) were inpatient cases, 75 (20.1%) were trauma/emergency cases, and eight (2.1%) were outpatient cases, and 229 of 373 (62.5%) were follow-up cases, of which 219 (95.6%) inpatient cases. Among the 144 new cases flagged for hemorrhage, 66 (44.4%) were positive, of which 39 (58.2%) were trauma/emergency cases. The overall sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 88.7%, 94.2% and 73.7%, 97.7%, and 93.4%, respectively. The accuracy of the intracranial hemorrhage detection was significantly higher for emergency cases than for inpatient cases (96.5% versus 89.4%). Conclusion This study reveals that the performance of the deep learning software for acute intracranial hemorrhage detection varies depending upon the patient visit location. Furthermore, a substantial portion of flagged cases were follow-up exams, the majority of which were inpatient exams. These findings can help optimize the artificial intelligence-driven clincical workflow.