1624
clinical study

Prescreening for Intracranial Hemorrhage on CT Head Scans with an AI-Based Radiology Workflow Triage Tool: An Accuracy Study

Materials & Methods

A retrospective dataset of 533 non-contrast head CT scans was collected from our large urban tertiary academic medical center. Following convention for studies evaluating sensitivities and specificities of imaging computer-aided detection and diagnosis devices, a prevalence-enriched dataset was utilized such that a 50% prevalence of intracranial hemorrhage was obtained. The algorithm was run on the dataset. Cases flagged by the algorithm as positive for ICH were defined as “positive”, and the rest as “negative.” The results were compared to the ground truth, determined by a neuroradiologist review of the dataset. 

Results

Algorithm sensitivity was 96.2% (CI: 93.2%-98.2%); specificity was 93.3% (CI: 89.6-96.0%). Estimated real-world NPV was determined as at least 96.2% (CI: 93.2%-97.9%), and an estimated upper threshold for PPV was estimated as 93.4% (CI: 90.1%-95.7%).

Conclusions

The tested device detects ICH with high sensitivity and specificity. The potential utility of using the device may be to autonomously surveil radiology worklists for studies containing critical findings and triage a busy workflow, ultimately improving patient care in clinically time-sensitive cases.

Explore the Latest AI Insights, Trends and Research