Assessing CT-based Volumetric Analysis via Transfer Learning with MRI and Manual Labels for Idiopathic Normal...

Area of Science:

• Radiology
• Artificial Intelligence
• Neurology

Background:

• Idiopathic normal pressure hydrocephalus (iNPH) diagnosis often relies on manual assessment of brain CT scans.
• Current methods for evaluating ventriculomegaly in iNPH involve visual rating and manual measurements, which can be time-consuming and subjective.
• Deep learning models offer potential for automated analysis of brain CT images.

Purpose of the Study:

• To enhance the segmentation of ventricular cerebrospinal fluid (VCSF) in brain CT scans using a deep learning model.
• To assess the performance of automated brain CT volumetrics in the diagnosis of iNPH patients.
• To evaluate the clinical utility of deep learning-derived CT-based volumetric measures (CTVMs) for iNPH assessment.

Main Methods:

• A two-stage deep learning approach was employed, initially training a 2D U-Net model on healthy controls and refining it with iNPH patient data.
• The model utilized a large dataset of CT scans from healthy controls and iNPH patients, with external validation datasets from multiple international centers.
• Three CT-based volumetric measures (CTVMs) relevant to iNPH were derived from the automated segmentation.

Main Results:

• Strong volumetric correlations (ϱ=0.91) were observed between automated and manual CT-VCSF measurements in iNPH patients.
• The CTVMs demonstrated high accuracy in differentiating iNPH patients from controls, with AUC values of 0.97 (external) and 0.99 (internal).
• The automated measures performed comparably to gold-standard assessments, even with intraventricular shunt catheters present.

Conclusions:

• Deep learning-derived CTVMs show significant potential for quantifying morphological features in hydrocephalus.
• Automated CT volumetrics can accurately distinguish iNPH patients from healthy controls, offering a valuable tool for diagnosis and monitoring.
• The widespread availability of CT makes this deep learning approach highly impactful for improving iNPH radiological evaluation globally.