Quantitative analysis of brain herniation from non-contrast CT images using deep learning.

Area of Science:

• Neurosurgery
• Radiology
• Artificial Intelligence

Background:

• Increased intracranial pressure (ICP) can lead to fatal brain herniation.
• Brain hematomas or tumors can cause mass effect, shifting the brain's midline.
• The midline shift (MLS) is a critical indicator of brain injury severity.

Purpose of the Study:

• To develop and validate an automated method for predicting midline shift (MLS) using a convolutional neural network (CNN).
• To assess the correlation between automated MLS estimation and clinical markers of brain hematoma severity.

Main Methods:

• A CNN algorithm was developed to predict the deformed midline (dML) in brain images.
• The algorithm was trained and validated on non-contrast computed tomography (NCCT) scans from 45 patients with epidural hemorrhage (EDH) and intra-parenchymal hemorrhage (IPH).

Main Results:

• The automated method accurately predicted MLS with low average errors in location (1.29 mm), 2D area (66.4 mm²), and 3D volume (253.73 mm³).
• Estimated MLS showed strong correlations with hematoma volume and a Radiologist-defined severity score (RSS).
• The automated system demonstrated excellent potential for clinical decision support in hematoma management.

Conclusions:

• The proposed automated CNN method provides accurate and efficient MLS estimation.
• This technology has significant clinical potential for improving hematoma severity assessment and decision support systems.
• Automated MLS prediction can aid clinicians in managing patients with brain hematomas.