Performance testing of a novel deep learning algorithm for the detection of intracranial hemorrhage and first trial under clinical conditions

Abstract

Purpose: We evaluate the performance of a deep learning-based pipeline using a Dense U-net architecture for detection of intracranial hemorrhage (ICH) in unenhanced head computed tomography (CT) scans. Methods: A balanced database was assembled retrospectively, comprising a total of 872 CT scans (362 with present ICH). Predictions by the algorithm were analyzed and compared to the radiology report (ground truth). Secondly, the algorithm's performance was tested in clinical environment: A total of 100 head CT scans (11 with present ICH) were analyzed simultaneously by the deep learning algorithm and a radiologist during clinical routine. The time until first temporary diagnosis of ICH was measured. Performances of the algorithm were evaluated in combination with the radiologist, when using it as triage tool. Results: In the retrospectively assembled dataset the deep learning algorithm detected ICH with a sensitivity of 91.4%, specificity of 90.4% and overall accuracy of 91.0%. In clinical environment, the algorithm was significantly faster compared to the temporary report of the assigned radiologist (24 ± 2 s vs. 613 ± 658 s, p < 0.001). When using the algorithm as a triage tool additional to the report of the assigned radiologist, a sensitivity of 100% was achieved. Conclusions: These results and the short processing time demonstrate the immense potential of deep learning applications for the use as triage tool and for additional review of manual reports.

Keywords

• Computed tomography
• Intracranial hemorrhage
• Artificial intelligence