AI learns to fix CT scans by protecting what algorithms actually measure

Computed tomography scans are usually judged by how good they look to a human eye, but a new study argues that this standard can actively work against automated biological measurement. Researchers have developed a projection-correction framework for CT data that optimizes images not for visual appeal, but for the downstream algorithms that extract measurements from them. The work, published in BMC Bioinformatics, uses feather morphometry as its motivating example and reports proof-of-concept results on both synthetic and real scan data.

The core insight is that acquisitions which appear overexposed or visually nonstandard may actually be appropriate when they expose subtle structures to an analysis algorithm. Traditional imaging pipelines optimize for human-readable images and generic quality metrics, yet algorithm-facing morphometric workflows require preservation of the specific information used by downstream measurement. The team therefore built a task-aligned framework in which a previously published feather-growth computer-vision workflow was treated as the downstream task to be protected during correction.

Technically, the method combines a residual denoising autoencoder with an expert human-preference ranker. The autoencoder defines a bounded projection-correction space, while the ranker scores candidate residual corrections for how well they preserve task-relevant morphology. Rank-weighted evolution-strategy updates then optimize latent corrections under projection-preservation, latent-bound, and angular-smoothing constraints, keeping the correction process within controlled limits rather than allowing arbitrary image transformations.

The benchmark consisted of five synthetic scenes: four development seeds and one fully held-out seed. Across the development-inclusive five-scene summary, the method achieved 67.9% lower reconstruction RMSE and 72.0% lower density-map RMSE than motion-corrupted data, with the fully held-out seed showing the same direction and a similar magnitude of improvement. Preference-guided correction also outperformed a feed-forward residual-autoencoder baseline, which recorded a reconstruction RMSE of 0.00260 ± 0.00013 and a density-map RMSE of 0.0500 ± 0.0021.

Comparison with an established tool was striking. TomoPy joint reprojection alignment changed the principal morphology-error endpoints by approximately 1% or less under the local non-rigid feather-motion benchmark, whereas the preference-guided approach produced substantial gains. Clean-data controls produced only small perturbations when no motion correction was needed, suggesting the method does not damage data that are already sound.

The framework was also tested in two illustrative real-data case studies. When inserted upstream of the published morphometric workflow, it reduced visible feather ghosting and improved downstream crest-density mapping and agreement with manual reference masks. These real-data results are presented as illustrative case studies rather than a large-scale evaluation, and the synthetic benchmark itself is a proof of concept with a small number of scenes.

The authors conclude that these results support task-aligned projection correction as a route toward CT imaging workflows optimized for downstream biological measurement rather than generic visual image quality alone. The study provides proof-of-concept evidence that expert preference signals can guide bounded projection-domain correction for surface-based biological morphometrics. Because the published version is an early-release article, it remains subject to further edits before the final Version of Record replaces it.

Subject of Research: A task-aligned projection-correction framework that uses expert preference signals to reduce CT motion artifacts in feather morphometry

Article Title: Preference-guided CT projection correction for surface-based feather morphometrics

Article References: Thompson, T., & Hsu, E. (2026). Preference-guided CT projection correction for surface-based feather morphometrics. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06716-5

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06716-5

Keywords: CT imaging, motion artifacts, feather morphometrics, projection correction, denoising autoencoder, preference learning, evolution strategies, TomoPy, computer vision, BMC Bioinformatics, open access, proof of concept