Accuracy of Artificial Intelligence-driven Automated Detection and Volumetric Assessment of intraventricular Hemorrhage

Introduction: The presence and increased volume of intraventricular hemorrhage (IVH) is associated with higher mortality and worse outcomes in patients with spontaneous intracerebral hemorrhage (ICH). Rapid diagnosis and volumetric assessment of IVH can therefore be used to help guide prognosis and determine appropriate patient care. Semi-Autonomous Segmentation (SAS) is the research gold standard for hemorrhage volume estimation, however it is not used real-time by clinicians due to the increased time for analysis. Recently, artificial intelligence (AI) software has been developed (Viz.ai, San Francisco, California) to automatically detect and volumetrically assess IVH. Our goal is to assess the accuracy of the AI segmentation algorithm as a tool for detecting IVH and determining IVH volume by comparing its performance to SAS.

Methods: Seventy-two head CTs positive for ICH were analyzed with SAS in 3D Slicer to determine IVH volume as the standard reference volume. The same CT scans were uploaded to the AI software for IVH detection and volume analysis.

Results: 28 (39%) of CTs were positive for IVH by a board-certified neurosurgeon. The AI algorithm had a sensitivity of 77.1%, specificity of 97.3%, positive predictive value of 96.4% and negative predictive value of 81.8%. False negatives have an average IVH volume of 1.1±0.7 mL and the false positive has an IVH volume of 1±0mL. The average difference in IVH volumetric assessment of true positives was 3.9±4.6 mL (r2=0.94).

Conclusion : This study indicates that AI offers an accurate estimation of ICH volume over a broad range of hematoma volumes when compared to standard SAS, which when coupled with significantly faster analysis compared to SAS justifies the use of AI in evaluation of IVH volume in the ICH triage workflow.