Automatic coarse-to-fine AC-PC localization on CT using registration-guided 3D-UNets

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Neuroimaging Analysis

Background:

• Accurate localization of the Anterior Commissure (AC) and Posterior Commissure (PC) is crucial for CT-based disease screening.
• Existing computational methods for AC-PC localization on CT scans are insufficient.
• This localization is foundational for calculating ventriculomegaly features used in diagnosing conditions like Normal Pressure Hydrocephalus (NPH).

Purpose of the Study:

• To develop and validate a novel registration-guided 3D-UNet framework for automated AC-PC localization on CT scans.
• To assess the framework's performance in computing ventriculomegaly features for NPH detection.
• To demonstrate the framework's utility in distinguishing NPH from other neurological conditions.

Main Methods:

• A registration-guided 3D-UNet framework was developed for CT-based AC-PC localization.
• The framework was trained and evaluated on internal (VA-Cohort, n=427) and external (VA-ExtCohort, UCSB-ExtCohort) datasets.
• Model development and evaluation used 1 mm³-resampled scans.

Main Results:

• The framework achieved low mean radial errors (MREs) across all cohorts, with upper 95% confidence intervals for localization errors below 3.2 mm.
• Ventriculomegaly features computed using the framework successfully distinguished NPH from Alzheimer's disease, post-traumatic volume loss, and headache (AUC=0.95).
• Performance closely matched that of manual AC-PC localization.

Conclusions:

• The registration-guided 3D-UNet framework provides accurate and automated AC-PC localization on CT scans, even with structural degeneration.
• This automated approach enables standardized radiological feature computation for neurodegenerative disease screening.
• The framework can enhance CT-based screening, particularly for elderly patients presenting with falls or cognitive changes.