Purpose: Early detection and diagnosis of intracranial aneurysms (IAs) are particularly critical. Deep learning models (DLMs) are now widely used in the diagnosis of various diseases. Different DLMs have been developed to detect IAs. However, the overall performance of various DLMs for detecting IAs has not been evaluated. We aimed at exploring the performance of DLMs in the detection of IAs and measuring the effect of DLMs in assisting clinicians.
Methods: A diagnostic accuracy meta-analysis using a mixed-effect binary regression model was performed to estimate accuracy in patient-level and lesion-level. Moreover, the effect in assisting clinicians was measured by a random-effect meta-analysis.
Results: Twenty cohort studies including a total of 17 DLMs were assessed eligible in the present study. We summarized that DLMs had both high sensitivity (0.92, 95 % CI: 0.85 to 0.96) and specificity (0.96, 95 % CI: 0.94 to 0.97) in the detection of IAs in patient-level. In lesion-level, we also found a high summary sensitivity of 0.92 (95 % CI: 0.87 to 0.95). Moreover, assisted by DLMs, the sensitivity of clinicians became higher (Risk ratio: 1.09, P-value: 0.0006), with no effect on the specificity in diagnosis (Risk ratio: 0.99, P-value: 0.53). The reading time was reduced by the assistance of the deep learning model. (Mean difference: -7.37, P-value: 0.0077) CONCLUSIONS: DLMs have the competence of detecting IAs accurately. Moreover, DLMs can improve clinicians' sensitivity and reduce the reading time without affecting the specificity in diagnosing IAs.
Keywords: Artificial intelligence; Deep learning; Intracranial aneurysm; Meta-analysis.
Copyright © 2022 Elsevier B.V. All rights reserved.