ACUS: An Angular Calibrated Uncertainty Score for Incomplete-View CT

Zihan Deng, Chuanzhi Xu, Quankeng Huang, Yujie Liu, Zhenzhen Gong, Junjie Huang, Tianye Niu, Lequan YuarXiv 2026 · arXiv

TL;DR

ACUS (Angular Calibrated Uncertainty Score) is a new metric designed to quantify exactly how much we should trust AI-generated CT scans when data is missing. By measuring the "angular gap" between a predicted view and the nearest actual X-ray measurement, ACUS provides a calibrated error bound that exposes the dangerous overconfidence of standard deep-learning models.

What Problem Does This Paper Address?

In modern medical imaging, we often perform "incomplete-view" CT scans—using fewer X-ray angles (sparse-view) or a restricted range of motion (limited-arc)—to reduce radiation dose or scan time. Deep learning models are remarkably good at "filling in the blanks," a process called projection completion.

However, there is a catch: these models rely on anatomical priors. They know what a human chest or jaw should look like, so they can "hallucinate" plausible-looking anatomy even where the scan didn't actually measure anything. Standard metrics like PSNR (Peak Signal-to-Noise Ratio) provide a single global score that can remain high even when the model is just making things up. This leaves clinicians with a critical question: How does the reliability of the image change as we move further away from the actual data points?

Figure 1. Incomplete-view CT as an interval problem. (a) Sparse- view and limited-arc scans leave angles with no ray coverage. A prior can complete them, and the question is whether the image is trustworthy. (b) PSNR, SSIM, and acquired-view geometry give no error bound on those angles. (c) ACUS reports width and coverage by the gap to the nearest acquired view: narrow nearby, wide far away. A static scale fails at large gaps.

Methods

The authors propose ACUS, which treats CT reconstruction as an interval estimation problem rather than just a point prediction.

  1. The Metric: ACUS is based on the Winkler interval score. It evaluates a nominal 90% predictive interval. The score rewards narrow intervals (high confidence) but applies a massive penalty if the ground-truth pixel falls outside that interval (a "miss").
  2. Angular Binning: Crucially, ACUS is binned by the angular gap ($\delta$)—the distance from the query angle to the nearest acquired measurement.
  3. The Architecture: The system uses a frozen Masked Autoencoder (MAE) backbone to encode acquired views. A "Scale Head" predicts pixel-wise uncertainty ($\sigma$) alongside the predictive mean ($\mu$).
  4. Consistency via Collapse: The authors introduce a "collapse loss." This forces the model's uncertainty to shrink to near-zero at the exact angles where data was acquired, serving as a necessary consistency check.
Figure 2. Interval estimator used to compute ACUS. (a) A frozen MAE encodes acquired views. Distance-weighted softmax fuses the two nearest into a mean µ and a scale σ. Point completion is the mean-only baseline. δ marks acquired, near, and far queries. (b) Chest and dental specialists train on sparse and multi-arc orbits with dropout and held-out queries. Gaussian likelihood has no variance target. Collapse drives σ toward zero and µ toward the acquired view. (c) The nominal 90% interval is µ ± 1.645σ (width 3.29σ). Inside, ACUS equals the width. Outside, it adds twenty times the distance to the bound.

Datasets

The researchers evaluated ACUS on two distinct anatomical domains:

  • LUNA16: A standard dataset for chest CT slices.
  • ToothFairy: A dataset for dental cone-beam CT (CBCT) volumes. Both were tested using simulated circular cone-beam orbits with sparse (8-view) and multi-arc sampling patterns.

Key Insights and Ideas

  • Uncertainty is Geometry-Dependent: The "aha moment" is that uncertainty isn't just about the image content; it's a function of the scan trajectory. As the angular gap $\delta$ increases, the "measurement fiber" (the set of all possible volumes consistent with the data) grows, and ACUS captures this expansion.
  • Point Estimates are Overconfident: Standard models often produce a high-quality "mean" image that looks correct but has zero "coverage"—meaning the true anatomy is frequently outside what the model's error bars should be.
  • The Static Scale Failure: Traditional baselines often use a fixed error estimate. ACUS shows that a static scale fails miserably at large gaps; you need a dynamic scale that "opens up" as you move away from known data.

Key Results

  • Monotonic Growth: In both chest and dental tests, the predicted interval width grew consistently as the angular gap increased. For chest scans, the mean scale rose from 0.018 (near) to 0.088 (far).
  • Calibrated Coverage: ACUS achieved empirical coverage between 77% and 92%, successfully approaching the 90% target.
  • Exposing Baselines: While point-completion models might have decent PSNR, ACUS revealed they had zero reported far coverage, meaning they were confidently wrong in the "blind spots" of the scan.
  • Visual Evidence: Figure 3 demonstrates that while standard models (NN, Point) show significant residuals (errors) in far-angle projections, ACUS maintains a "sparse miss mask," indicating its predicted intervals actually contain the truth more often.
Figure 3. Far-angle held-out projections on chest and dental cone-beam orbits. Columns: ground truth (GT), nearest neighbor (NN), point completion, ACUS mean, ACUS scale, color residual maps for NN, point completion, and ACUS, and miss masks, with a shared residual colorbar. Constant-scale baselines miss most far pixels. ACUS keeps a structured scale and a sparse miss mask.

Limitations and Future Work

  • Pixel Independence: The current model assumes pixel-wise independence in its Gaussian distribution. In reality, errors in CT reconstructions are spatially correlated.
  • Simulated vs. Real: The study uses simulated noise-free projections. Real-world scans involve photon noise, scatter, and patient motion, which would likely broaden the required uncertainty intervals.
  • Calibration Guarantees: While the empirical coverage is good, the authors note that ACUS is a diagnostic tool, not yet a formal statistical "operating guarantee." Future work could involve integrating conformal prediction for more rigorous bounds.

Summary updated

Abstract

Sparse-view and limited-arc computed tomography (CT) are commonly treated as projection completion: a network fills an unmeasured view, and the result is ranked by similarity to one ground-truth image. That score can be high even when the query angle is far from every acquired measurement, because a learned prior may invent anatomy the scan never constrained. Existing protocols therefore leave a basic question unanswered: how does predictive reliability change with distance from the acquired views? We propose the Angular Calibrated Uncertainty Score (ACUS) to answer that question. ACUS is the Winkler interval score of a nominal 90% predictive interval on held-out projections, binned by distance to the nearest acquired view. It charges interval width and adds a large miss penalty when a held-out pixel falls outside the interval, so a narrow band helps only when coverage holds. We use collapse on acquired angles as a necessary consistency check on the predictor. On chest and dental cone-beam orbits, interval width grows with the angular gap. Empirical coverage ranges from 77% to 92%, and the peak signal-to-noise ratio (PSNR) of the point estimate falls as the width rises. Nearest-neighbor, interpolation, and point completion, each given one residual scale, have zero reported far coverage and incur outside-distance penalties. ACUS exposes gap-dependent overconfidence that a point-estimate PSNR alone does not measure.
BibTeX
@misc{deng2026acus,
  author = {Zihan Deng and Chuanzhi Xu and Quankeng Huang and Yujie Liu and Zhenzhen Gong and Junjie Huang and Tianye Niu and Lequan Yu},
  title = {{ACUS: An Angular Calibrated Uncertainty Score for Incomplete-View CT}},
  year = {2026},
  eprint = {2611.23752},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23752},
}

Loading related content...

Figure