Normal Pressure Hydrocephalus Classification using Weakly-Supervised Local Feature Extraction.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Normal Pressure Hydrocephalus (NPH) diagnosis is challenging due to overlapping symptoms with other neurological disorders.
• Ventricular cerebrospinal fluid (CSF) volume is a key radiological indicator, but expert assessment is limited.
• Existing automated tools struggle with generalization and often overlook local CSF volume variations.

Purpose of the Study:

• To develop a weak supervision method for training CSF segmentation models without expert annotations.
• To introduce a local volumetric feature extraction algorithm for enhanced NPH classification.
• To improve automated screening of patients at risk for NPH.

Main Methods:

• A novel weak supervision technique was employed to train a CSF segmentation model from scratch on a target dataset.
• A local volumetric feature extraction algorithm was developed to capture regional CSF volume differences.
• The combined approach was evaluated on non-contrast CT scans of 105 NPH and 112 non-NPH patients.

Main Results:

• The proposed method achieved high NPH classification performance: ACC = 0.88, Sen = 0.97, Spec = 0.79, F1 = 0.89, AU-ROC = 0.91.
• Outperformed existing segmentation methods in NPH classification accuracy.
• Demonstrated superior screening performance compared to neuroradiologists' visual evaluations.

Conclusions:

• Weakly-supervised CSF segmentation combined with local volumetric features offers a robust solution for NPH diagnosis.
• The developed automated tool can significantly aid in the early and accurate screening of NPH patients.
• This approach addresses the limitations of current automated methods by avoiding reliance on extensive prior knowledge and incorporating local volumetric data.