The stakes are larger than they might first appear. Normal pressure hydrocephalus, a condition in which excess cerebrospinal fluid distends the ventricles and produces a treatable triad of gait disturbance, cognitive decline, and urinary incontinence, is misdiagnosed as irreversible dementia in an estimated 80 percent of cases. The disease is surgically reversible with a shunt, yet distinguishing it from Alzheimer’s disease and other atrophic mimics requires precise, standardized measurements of ventricular enlargement, measurements that are only valid when computed in reference to the AC-PC line. Without accurate localization of those two commissures, the diagnostic features that guidelines recommend, such as the Evans Index and the callosal angle, lose their reliability and their discriminative power.

The problem is compounded by the realities of clinical practice. Falls cause roughly three million emergency department visits in the United States every year, and CT is the primary modality used to evaluate them. It is faster, cheaper, more widely available, less prone to geometric distortion, and safer for patients with implants than MRI. An analysis of a decade of records from a national Veterans Affairs database, described in the study, found that among patients with suspected dementia and unlikely head injury, approximately 76 percent received CT only, 10 percent MRI only, and 14 percent both. In other words, the scans most likely to reveal underlying neurological disease in this vulnerable population are precisely the scans on which robust automated AC-PC localization has been lacking.

Existing computational tools offer little help. The only dedicated solution for AC-PC alignment on brain CT, a registration-based tool called fatbACPC, relies on iteratively matching intensity distributions between a patient scan and a template. Registration cost functions such as normalized correlation or mutual information depend on clear contrast boundaries between adjacent structures, yet the internal brain parenchyma on CT lacks the distinct soft-tissue contrast that drives accurate local alignment. Non-linear registration can better accommodate globally warped pathological anatomy than rigid approximations, but even the best registration struggles to map severely distended ventricles in hydrocephalus or cortical volume loss in atrophy onto a normal template. The Minnesota team hypothesized that a deep learning model informed by registration could overcome these limitations through a coarse-to-fine strategy borrowed from low-contrast cephalometric landmark prediction.

The framework works in two stages. In the coarse stage, each skull-stripped CT scan is non-linearly registered to a CT template of the elderly brain using B-spline parameterized transformations optimized with mutual information, and known template coordinates of the AC and PC are mapped back into the native scan space to yield initial estimates. In the fine stage, a modified 3D-UNet, the volumetric cousin of the architecture that revolutionized biomedical image segmentation, receives two-channel inputs comprising 32-millimeter isotropic patches cropped around the coarse estimates. The network is trained to predict Gaussian confidence heatmaps for the AC, the PC, and a normalizing background, using a loss function that combines a focal loss for the dense foreground landmarks with cross-entropy for the over-represented background. Final landmark coordinates are inferred as the mean location of active voxels within each output patch.

Because routine CT natively supports axial, coronal, and sagittal reformats, the researchers trained reformat-specific models sharing identical architecture and hyperparameter tuning, then integrated the predictions: the x-coordinate averaged from axial and coronal models, the y-coordinate from axial and sagittal models, and the z-coordinate from coronal and sagittal models. This design deliberately excludes the sub-optimally resolved slice dimension of each reformat and aligns with the multi-planar nature of the downstream diagnostic features. All scans passed through an automated preprocessing pipeline of gantry tilt correction, isotropic resampling to one-cubic-millimeter voxels with cubic B-splines, cropping, bed and skull removal, and windowing, with a zero percent failure rate confirmed by visual inspection.

Reference standards were established under the guidance of a board-certified neuroradiologist with 34 years of experience. A primary annotator trained by this expert labeled the internal cohort of 427 scans from 340 patients with normal pressure hydrocephalus, Alzheimer’s disease, post-traumatic volume loss, and headache, while three blinded secondary annotators provided independent annotations on mutually exclusive subsets to verify inter-annotator agreement. On 98.5 percent of scans, the primary and secondary annotations agreed within 3.2 millimeters, and an extended Bland-Altman analysis on a common set of 36 scans showed no annotator bias with agreement limits within 1.1 millimeters across all coordinates. That 3.2-millimeter figure would prove central to the entire study.

Under five-fold cross-validation on the internal cohort, the framework achieved mean radial errors of 1.64 millimeters for the AC and 1.49 millimeters for the PC, with the upper bounds of the 95 percent confidence intervals well below the 3.2-millimeter threshold. External validation on two independent datasets, 40 multi-planar scans from VA patients with ventricular shunting for hydrocephalus of various etiologies and 43 axial-only scans from the University of California, Santa Barbara, yielded mean radial errors of 2.42 and 1.79 millimeters and 2.31 and 1.93 millimeters respectively, with landmark predictions within 3.2 millimeters of the reference standard in more than 83 percent of scans. The benchmarks were emphatic: a baseline 3D-UNet performing single-step localization across whole scans without registration guidance produced mean radial errors of 9.46 and 9.95 millimeters on the internal cohort, while fatbACPC reached 3.93 and 5.09 millimeters on axial reformats, markedly worse than the fine-stage refinement. The researchers empirically established the clinical significance of the 3.2-millimeter limit through a perturbation study in which artificially introduced localization noise significantly degraded the discriminative power of AC-PC-referenced ventriculomegaly features beyond that threshold.

The clinical payoff came in the classification study. On a chart-verified subset of 238 scans, ventriculomegaly features computed using the framework’s automated AC-PC predictions distinguished normal pressure hydrocephalus from Alzheimer’s disease, post-traumatic volume loss, and headache with a test-set area under the receiver operating characteristic curve of 0.95, an accuracy of 89 percent, a sensitivity of 92 percent, and a specificity of 87 percent. McNemar’s tests confirmed this performance showed no significant difference from models trained on features standardized with manually annotated ground-truth AC-PC coordinates, across both multi-class and binary settings. Individual feature values differed negligibly between the two approaches, and univariate analysis verified that classification was not driven by circularity bias from the Evans Index diagnostic criterion. The standalone predictive power of the AC-PC-independent cerebrospinal fluid to brain volume ratio, with areas under the curve of 0.78 and 0.90, was substantially lower than the combined feature set, reinforcing the necessity of accurate commissure localization.

Inference is fast enough for the clinic. Preprocessing and coarse localization required about 122 seconds per scan, while the three reformat-specific 3D-UNet inferences each took roughly a third of a second on a consumer GPU, reducing total processing to under 42 seconds per scan if reformats are processed in parallel. The framework proved robust to head tilt, native slice thickness, and patient age, none of which significantly affected localization errors, and comparisons between male and female subjects showed no significant sex biases. The authors note limitations, including a predominantly male veteran sample, small external test-only datasets, and registration failures on scans with extreme head tilts beyond 40 degrees, which they say necessitate radiologist oversight for such outliers. They emphasize that real-world clinical integration should include clinician verification of localized landmarks and downstream feature computation. Once validated in larger, sex-balanced external cohorts, the tool could support clinical decision-making in high-volume settings like emergency departments, potentially avoiding unnecessary MRIs by maximizing the diagnostic utility of the quicker, more accessible CT, and offering a standardized computational pathway for screening neurodegenerative diseases characterized by the interplay of hydrocephalus and atrophy.

Subject of Research: Automatic AC-PC landmark localization on CT using registration-guided 3D-UNets for normal pressure hydrocephalus screening

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

Article References: Kadaba Sridhar, S., Eastman, A., Wilson, P., Mishra, S., Broadbent, C., Truwit, C., Kuang, R., & Samadani, U. (2026). Automatic coarse-to-fine AC-PC localization on CT using registration-guided 3D-UNets. Frontiers in Neuroinformatics, 20, Article 1853412. https://doi.org/10.3389/fninf.2026.1853412

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1853412

Keywords: anterior commissure, posterior commissure, computed tomography, deep learning, 3D-UNet, image registration, normal pressure hydrocephalus, ventriculomegaly, Alzheimer's disease, neurodegenerative disease screening, brain imaging, artificial intelligence