Deep learning-based prediction of mortality using brain midline shift and clinical information.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Brain midline shift (MLS) is a critical indicator of intracranial mass effect severity, essential for emergency surgical decisions and prognosis.
• Immediate diagnosis of MLS is crucial due to its emergent nature.
• Current 2D CT slice analysis poses limitations for accurately assessing the 3D midline structure.

Purpose of the Study:

• To develop a computer-aided deep learning method for detecting brain MLS on CT slices.
• To predict patient mortality by combining MLS measurements with clinical information.
• To improve the accuracy and efficiency of MLS detection and its prognostic value.

Main Methods:

• A keypoint detection deep learning model was proposed to identify the brain midline on individual CT slices.
• MLS distance and area were quantified from each slice.
• A multilayer perceptron (MLP) model integrated MLS data with clinical information for mortality prediction.

Main Results:

• The model achieved high accuracy (0.966), precision (0.952), sensitivity (0.991), specificity (0.932), and F1-score (0.971) in slice selection and MLS detection.
• MLS distance and volume were predicted with minimal error at both slice and case levels.
• The MLP model demonstrated good accuracy (above 0.8) and AUC in predicting patient mortality, particularly for cases with large MLS.

Conclusions:

• The proposed deep learning method effectively detects brain midline shift on CT scans.
• Integrating MLS measurements with clinical data using an MLP model provides a robust approach for predicting patient prognosis.
• This AI-driven tool shows promise for immediate diagnosis and improved outcomes in patients with intracranial lesions.