An emerging deep learning software may offer enhanced detection of intracranial hemorrhages on non-contrast cranial computed tomography (cCT) scans, according to new research.
For the retrospective study, recently published in Insights into Imaging, researchers reviewed data from 2,960 cranial CTs for 2,486 patients (median age of 74) to assess the adjunctive capability of the AI software syngo.CT Brain Hemorrhage VB60 (Siemens Healthineers) to detect intracranial hemorrhage (ICH). The cohort included 314 cranial CTs that were positive for ICH, according to the study.
Overall, the researchers found that the AI software provided sensitivity and specificity rates of 93.6 percent and 96.2 percent respectively. The study authors also pointed out that the AI software had an accuracy rate of 95.9 percent and a negative predictive value (NPV) of 99.2 percent.
“The high NPV of the extended prototype (99.2%) underscores the high reliability of its negative predictions, indicating that the AI can effectively preselect cases and flag almost all conspicuous findings with a low false negative rate of 7 cases per 1000. This result suggests potential for workload reduction, particularly during periods of limited staffing, and to help prioritize clinically relevant cases,” noted lead study author Franziska Tombach, M.D., who is affiliated with the Department of Diagnostic and Interventional Radiology at the University Hospital of Wurzburg in Wurzburg, Germany, and colleagues.
The study authors also observed that adjunctive use of the AI software detected 12 cases of ICH that were missed in initial review by radiologists.
“This underscores the efficacy of concurrent AI integration during routine reporting. … The concomitant application of AI alongside radiologists may augment diagnostic confidence. In instances of conflicting interpretations, this dual reading approach fosters critical analysis and, when necessary, the rectification of misinterpretations,” added Tombach and colleagues.
Three Key Takeaways
• High diagnostic performance for ICH detection. The AI software demonstrated strong accuracy (95.9 percent) with high sensitivity (93.6 percent) and specificity (96.2 percent), and an excellent negative predictive value (99.2 percent), supporting reliable exclusion of hemorrhage on non-contrast head CT.
• Clinical workflow augmentation and error reduction. Adjunctive AI use identified additional ICH cases missed on initial radiologist review (12 cases) and may enable effective triage, workload reduction, and prioritization, which is particularly valuable in high-volume or understaffed settings.
• Encouraging performance in hemorrhage subtyping, albeit with reduced specificity. The AI demonstrated reasonable sensitivity for detecting subarachnoid hemorrhage (SAH) (87.1 percent), supporting its potential role in early identification and triage for clinically significant subtypes. However, the researchers also noted reduced specificity (67.9%) for SAH, reinforcing the need for radiologist confirmation, particularly when downstream management (e.g., CTA) is impacted.
In a subtype analysis, the researchers found that the AI software provided an 87.1 percent sensitivity, 67.9 percent specificity and a 74.5 percent NPV for subarachnoid hemorrhage (SAH).
“ … The detection and classification of SAH is of particular relevance, as atraumatic SAH is often associated with aneurysm rupture and typically necessitates additional CT angiography. This principle similarly applies to atypical parenchymal hemorrhages, necessitating further investigation. Accurate subtype classification has the potential to identify such findings in a timely manner, enabling prompt follow-up examinations,” emphasized Tombach and colleagues.
(Editor’s note: For related content, see “FDA Clears AI-Powered CT Triage Software for Intracranial Hemorrhage,” “Can Radiomics Enhance Differentiation of Intracranial Aneurysms on Computed Tomography Angiography?” and “Study: Deep Learning Denoising May Facilitate Up to a 75 Percent Reduction in Radiation Dosing for Head CT.”)
Beyond the inherent limitations of a single-center retrospective study, the authors acknowledged the use of a single CT scanner model and a lack of assessment for integration of the AI model into the clinical workflow.
