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

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Radiology

Background:

• Machine learning, particularly deep learning, is revolutionizing medical image analysis.
• Convolutional Neural Networks (CNNs) have demonstrated significant potential in various neuroimaging applications.
• Accurate detection and quantification of subdural hematoma (SDH) are crucial for effective patient management.

Purpose of the Study:

• To evaluate the performance of a novel CNN for detecting subdural hematoma (SDH) on noncontrast head CT (NCHCT).
• To assess the CNN's ability to quantify SDH thickness, volume, and midline shift (MLS).
• To validate the CNN's performance on an independent dataset.

Main Methods:

• Retrospective analysis of NCHCT studies from 263 patients evaluated for head trauma.
• Ground truth for SDH, thickness, and MLS established by neuroradiology reports.
• External validation of the CNN using area under the receiver operating characteristic curve (AUC) analysis for detection and mean absolute error for quantification.

Main Results:

• The CNN achieved high accuracy in detecting SDH, with an overall accuracy of 95.1% (95% CI 91.7%-97.3%).
• Sensitivity and specificity for SDH detection were 91.4% and 96.4%, respectively.
• Excellent agreement was found between automated and manual measurements for thickness (Pearson correlation coefficient 0.97) and MLS.

Conclusions:

• The evaluated CNN demonstrates exceptional performance in identifying and quantifying key features of SDH from NCHCT.
• The AI tool shows high accuracy for measuring SDH thickness and midline shift.
• This CNN holds promise for improving the efficiency and accuracy of SDH assessment in clinical practice.