Investigation of probability maps in deep-learning-based brain ventricle parcellation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Normal Pressure Hydrocephalus (NPH) is a neurological disorder characterized by ventriculomegaly.
• Accurate segmentation of the brain's ventricular system is essential for NPH diagnosis and surgical planning.
• Existing segmentation methods may struggle with enlarged ventricles or post-surgical artifacts in MRI.

Purpose of the Study:

• To develop and validate a modified 3D U-net model for precise ventricle parcellation in MRI scans.
• To enhance the accuracy of segmentation for NPH patients, including those with significant ventricular enlargement or shunt artifacts.
• To provide a robust tool for evaluating the ventricular system in the context of NPH management.

Main Methods:

• Modification of a 3D U-net architecture incorporating probability maps.
• Application of the modified 3D U-net for ventricle segmentation on brain MRIs.
• Evaluation of segmentation performance using Dice Similarity Coefficient (DSC) on healthy controls and NPH patients.

Main Results:

• Achieved high mean DSC values for whole ventricle segmentation: 0.864 ± 0.047 (healthy controls) and 0.961 ± 0.024 (NPH patients).
• Demonstrated superior performance on MRIs with grossly enlarged ventricles (mean DSC: 0.965 ± 0.027) and post-surgery shunt artifacts (mean DSC: 0.964 ± 0.031).
• The proposed method shows robustness and comparable performance to state-of-the-art techniques.

Conclusions:

• The modified 3D U-net with probability maps offers a robust and accurate tool for brain ventricle parcellation.
• This method effectively handles challenges like enlarged ventricles and surgical artifacts in NPH patient MRIs.
• The developed technique supports improved evaluation for NPH surgical intervention.