AI-Based Approach to One-Click Chronic Subdural Hematoma Segmentation Using Computed Tomography Images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Chronic subdural hematoma (CSH) is a condition with increasing incidence in older adults.
• Manual segmentation of CSH from CT scans is time-consuming and requires expertise.
• Existing segmentation methods may lack efficiency and clinical integration.

Purpose of the Study:

• To develop and evaluate a one-click, computer vision-based automated segmentation tool for chronic subdural hematomas.
• To compare the performance of the automated tool against manual segmentation by medical experts.
• To assess the clinical utility and efficiency of the proposed segmentation approach.

Main Methods:

• Development of a custom dataset comprising 53 CT scan series from 21 patients.
• Training of two U-Net based neural network models for automated CSH segmentation.
• Utilizing 10-fold cross-validation and the Dice metric for performance evaluation.
• Comparison with manual segmentation by three medical experts on an independent test set.

Main Results:

• The best performing model achieved a Dice score of 0.77 for segmentation accuracy.
• The one-click automated segmentation was over seven times faster than manual segmentation.
• Segmentation quality was deemed acceptable for clinical use by medical experts.

Conclusions:

• A novel one-click computer vision approach enables rapid and accurate segmentation of chronic subdural hematomas.
• The developed tool, integrated as an OsiriX plugin, offers a significant time advantage over manual methods.
• This automated segmentation shows potential for improving clinical workflow efficiency in neuroimaging.