Automated detection and analysis of subdural hematomas using a machine learning algorithm

Abstract

OBJECTIVE Machine learning algorithms have shown groundbreaking results in neuroimaging. Herein, the authors evaluate the performance of a newly developed convolutional neural network (CNN) to detect and quantify the thickness, volume, and midline shift (MLS) of subdural hematoma (SDH) from noncontrast head CT (NCHCT). METHODS NCHCT studies performed for the evaluation of head trauma in consecutive patients between July 2018 and April 2021 at a single institution were retrospectively identified. Ground truth determination of SDH, thickness, and MLS was established by the neuroradiology report. The primary outcome was performance of the CNN in detecting SDH in an external validation set, as measured using area under the receiver operating characteristic curve analysis. Secondary outcomes included accuracy for thickness, volume, and MLS. RESULTS Among 263 cases with valid NCHCT according to the study criteria, 135 patients (51%) were male, the mean (± standard deviation) age was 61 ± 23 years, and 70 patients were diagnosed with SDH on neuroradiologist evaluation. The median SDH thickness was 11 mm (IQR 6 mm), and 16 patients had a median MLS of 5 mm (IQR 2.25 mm). In the independent data set, the CNN performed well, with sensitivity of 91.4% (95% CI 82.3%–96.8%), specificity of 96.4% (95% CI 92.7%–98.5%), and accuracy of 95.1% (95% CI 91.7%–97.3%); sensitivity for the subgroup with an SDH thickness above 10 mm was 100%. The maximum thickness mean absolute error was 2.75 mm (95% CI 2.14–3.37 mm), whereas the MLS mean absolute error was 0.93 mm (95% CI 0.55–1.31 mm). The Pearson correlation coefficient computed to determine agreement between automated and manual segmentation measurements was 0.97 (95% CI 0.96–0.98). CONCLUSIONS The described Viz.ai SDH CNN performed exceptionally well at identifying and quantifying key features of SDHs in an independent validation imaging data set.

Advances in chronic subdural hematoma and membrane imaging

“…Given the heterogeneity of cSDH morphology, geometry, and location, volumetric analysis likely provides more accurate information regarding cSDH size; however, imaging segmentation of cSDHs can be time- and resource-intensive and is not routinely employed in clinical practice. Artificial intelligence applications are available ( 25 ), and they may be employed to allow for rapid quantification of cSDH volumes in clinical use.…”

Machine Learning in Neuroimaging of Traumatic Brain Injury: Current Landscape, Research Gaps, and Future Directions

“…Chen (2022) employed a CNN for volumetric assessment, with the results closely mirroring manual segmentation with an AUROC of 0.83 [89]. Another CNN model designed for comprehensive SDH evaluation, including thickness, volume, and midline shift, showcased a sensitivity of 91.4% and specificity of 96.4% [90].…”

Artificial Intelligence in Neuroradiology: A Review of Current Topics and Competition Challenges

“…False positives were associated with the observation of calcifications and other hyperdense structures on CT. Colasurdo et al used the Viz.ai software package to detect subdural hemorrhage (SDH). The package had a sensitivity and specificity of 91% and 96%, respectively, with sensitivities smaller for small chronic SDH [ 54 ]. Another study, by Seyam et al, used Aldoc Medical to detect subtypes of hemorrhage.…”