Diagnosis of intracranial hemorrhages based on brain computed tomography with artificial intelligence

Abstract

BACKGROUND: Intracranial hemorrhages are associated with high mortality and risk of disability, requiring prompt and accurate diagnosis, particularly within the first 24 hours. The use of artificial intelligence technologies in analyzing brain computed tomography scans can shorten diagnostic time and improve diagnostic quality. The relevance of this study is emphasized by the limited number of certified artificial intelligence services for detecting intracranial hemorrhages in Russia and lacking data on their long-term effectiveness, highlighting the need for multicenter monitoring to assess the stability and accuracy of such systems in clinical practice. AIM: The study aimed to assess the diagnostic accuracy and stability of an artificial intelligence service in detecting intracranial hemorrhages on non-contrast brain computed tomography scans in a multicenter clinical monitoring setting for 18 months. METHODS: Anonymized brain computed tomography scans were used. The artificial intelligence service underwent a three-phase evaluation to evaluate its diagnostic accuracy and clinical performance using limited datasets. Two radiologists specializing in neuroimaging examined 80 brain computed tomography scans each month for 18 months, which had been preprocessed by the artificial intelligence service and randomly selected from the clinical workflow. The results were analyzed using ROC analysis with sensitivity, specificity, accuracy, and area under the curve. RESULTS: During clinical monitoring, 1200 brain computed tomography scans were analyzed, with signs of intracranial hemorrhage detected in 48.3% of the scans. Based on the binary classification of intracranial hemorrhage presence or absence performed by the artificial intelligence service, the following diagnostic metrics were obtained: sensitivity, 97.4% (95.8–98.5); specificity, 75.4% (71.8–78.7); accuracy, 86.0% (83.9–87.9); and area under the curve, 94% (92.6–95.3). Eventually, a significant moderate positive correlation was observed in most diagnostic metrics and the time variable, except for sensitivity, which was affected by an update to the service version. However, full concordance between artificial intelligence-based markings and radiologist conclusions was noted in 28.5% of cases of identified intracranial hemorrhage, whereas discrepancies were found in 71.5%. The refined diagnostic metrics for cases with complete agreement with the radiologists’ report were as follows: sensitivity, 26.6%; specificity, 73.8%; accuracy, 50.1%; and area under the curve, 49.6%. CONCLUSION: The current configuration of the artificial intelligence service allows ruling out intracranial hemorrhage with very high probability, which may be useful in the initial triaging of patients in emergency settings. However, low values of refined metrics indicate considerable discrepancies between radiologist reports and service-generated results regarding the interpretation of pathological findings.

Ähnliche Arbeiten