An Artificial Intelligence Algorithm Integrated into the Clinical Workflow Can Ensure High Quality Acute Intracranial Hemorrhage CT Diagnostic.

Abstract

Purpose: Intracranial hemorrhage (ICH) is a life-threatening condition requiring rapid diagnostic and therapeutic action. This study evaluates whether Artificial intelligence (AI) can provide high-quality ICH diagnostics and turnaround times suitable for routine radiological practice.

Methods: A convolutional neural network (CNN) was trained and validated to detect ICHs on DICOM images of cranial CT (CCT) scans, utilizing about 674,000 individually labeled slices. The CNN was then incorporated into a commercial AI engine and seamlessly integrated into three pilot centers in Germany. A real-world test-dataset was extracted and manually annotated by two experienced experts. The performance of the AI algorithm against the two raters was assessed and compared to the inter-rater agreement. The overall time ranging from data acquisition to the delivery of the AI results was analyzed.

Results: Out of 6284 CCT examinations acquired in three different centers, 947 (15%) had ICH. Breakdowns of hemorrhage types included 8% intraparenchymal, 3% intraventricular, 6% subarachnoidal, 7% subdural, < 1% epidural hematomas. Comparing the AI's performance on a subset of 255 patients with two expert raters, it achieved a sensitivity of 0.90, a specificity of 0.96, an accuracy of 0.96. The corresponding inter-rater agreement was 0.84, 0.98, and 0.96. The overall median processing times for the three centers were 9, 11, and 12 min, respectively.

Conclusion: We showed that an AI algorithm for the automatic detection of ICHs can be seamlessly integrated into clinical workflows with minimal turnaround time. The accuracy was on par with radiology experts, making the system suitable for routine clinical use.