🧬 biology

Preference Guided CT Projection Correction for Surface-Based Feather Morphometrics

This paper presents a preference-guided CT projection correction framework that utilizes expert human ranking to optimize motion artifact reduction for downstream biological morphometrics, achieving significant improvements in reconstruction and density-map accuracy over traditional visual denoising methods.

Original authors: Tyler Thompson, Edward Hsu

Published 2026-07-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tyler Thompson, Edward Hsu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to solve a mystery by looking at a 3D hologram of a crime scene. Usually, when we take pictures or scans, we want them to look pretty and clear to our human eyes. We want the colors to pop and the lines to be sharp. But what if the person you are trying to help isn't a human detective, but a super-smart computer program? This program doesn't care if the picture looks "nice"; it only cares if it can count tiny objects accurately. In the world of medical and biological imaging, scientists use a machine called a CT scanner (think of it like a super-powered, 3D X-ray camera) to take pictures of tiny things, like bird feathers. Sometimes, the bird twitches or the feathers flutter while the machine is snapping photos. To a human, the resulting blurry picture might just look a little "fuzzy." But to the computer trying to count those feathers, that fuzziness looks like extra feathers or fake ridges. If the computer counts these fake feathers, the whole study is wrong. So, the big question is: how do we fix a blurry picture not to make it look better for us, but to make it work better for the computer?

This paper introduces a clever new way to fix those blurry CT scans, specifically for studying bird feathers. The researchers, Tyler Thompson and Edward Hsu from the University of Utah, built a system that acts like a "smart editor" for these scans. Instead of asking a human to say, "Does this look clean?", they taught a computer to ask, "Does this look like it will help the feather-counting program do its job?" They used a method called "preference learning," where a human expert looked at pairs of slightly different versions of a blurry feather image and picked the one that seemed less likely to trick the computer. The computer learned from these choices to find the perfect fix.

Here is the magic trick they used: they didn't try to redraw the whole picture from scratch. Instead, they used a "residual autoencoder," which is like a very smart guesser that only suggests tiny, safe changes to the blurry parts. Then, they used a "preference ranker" (the computer that learned from the human expert) to guide a search through millions of tiny possible changes. It's like having a blindfolded hiker trying to find the highest peak in a foggy mountain range. Instead of walking randomly, the hiker has a guide who says, "That step feels right; keep going that way," based on what the guide knows about the terrain. In this case, the "terrain" is the world of accurate feather counting.

The team tested this on a digital simulation of a pigeon with 300 feathers. They simulated the feathers fluttering wildly (up to 3 degrees of motion) to create "ghosting" effects, where the computer sees double or triple the feathers. When they applied their new "preference-guided" fix, the results were impressive. The error in the reconstructed 3D image dropped by about 68%, and the error in the feather density maps (the computer's count of where feathers are) dropped by a huge 72%. This was much better than just using standard "visual" cleaning tools, which only reduced the error by about 20-25%. The researchers also checked that their method didn't accidentally mess up pictures that were already clear; when they ran it on clean data, the changes were tiny and harmless.

To prove this wasn't just a computer game, they tried it on two real pigeon scans. In the real world, the feathers were indeed "ghosting" (showing up as double edges). After their correction, the "ghosts" disappeared, and the computer's map of the feathers matched a human expert's drawing much more closely. The agreement between the computer's mask and the human's mask jumped from about 74% to 89% in one bird and from 78% to 91% in the other.

The paper suggests that this approach is a game-changer for biological imaging. It argues that we shouldn't just fix images to make them look pretty for humans; we should fix them to make them useful for the specific task at hand. While the study is a "proof-of-concept" (meaning it's a successful first step rather than a finished product ready for every hospital), it shows that using human expert preferences to guide computer corrections can lead to much more accurate scientific measurements. The researchers admit their simulation was simplified and they only tested it on two real birds, so more work is needed, but the core idea—that we should tune our image fixes for the task, not just the eye—seems to hold up very well.

Technical Summary: Preference Guided CT Projection Correction for Surface-Based Feather Morphometrics

Problem Statement
Traditional CT imaging pipelines often optimize for human-readable image quality or generic metrics (e.g., sharpness, noise reduction). However, for quantitative biological morphometrics, the critical endpoint is the accuracy of downstream measurements derived from reconstructed surfaces, such as feature counts, density maps, and shape analysis. In CT scans of feathered specimens, small within-view motions of low-opacity feathers (feather flutter) can create subtle edge duplications and ghosting in projection data. While these artifacts may appear modest to a human observer, they propagate nonlinearly through reconstruction and surface extraction, resulting in structured geometric errors. These errors manifest as duplicated surface ridges, inflated component counts, and biased density maps, leading to incorrect biological measurements. Conventional correction methods evaluated on visual fidelity do not necessarily preserve the specific information required for accurate morphometric analysis.

Methodology
The authors propose a task-aligned projection-correction framework that optimizes for downstream biological measurement rather than generic visual quality. The workflow consists of the following components:

Synthetic Benchmark: A proof-of-concept benchmark was established using five independently seeded synthetic scenes based on a pigeon body-derived surface model. Synthetic feathers were placed on the surface and subjected to simulated within-view "flutter" motion (severe 3° amplitude) while the body remained fixed. The simulation included realistic detector noise, gain variations, and stripe artifacts.
Residual Autoencoder: A residual denoising autoencoder defines a bounded latent space for projection corrections. Instead of replacing the measured projection, the network proposes a bounded residual correction field ( $\tilde{x} = x + \text{clip}[D_\psi(z), -r_{max}, r_{max}]$ ). This ensures the correction remains local and does not hallucinate specimen structures.
Human-Preference Ranker: A Siamese ResNet-18 architecture was trained on pairwise preferences from a single expert. The ranker evaluates candidate corrected projection patches and predicts which is less likely to produce downstream surface artifacts (feather ghosting/duplication). This replaces scalar motion amplitude as the supervision signal, allowing the system to learn a morphology-risk function over continuous candidate corrections.
Preference-Guided Optimization: An evolution-strategy (ES) approach navigates the autoencoder's latent space. Candidate latent vectors are generated via Gaussian perturbations, decoded into residuals, and scored by the preference ranker. The ranking scores are converted into rank weights to estimate a gradient-free update direction. The optimization is constrained by:
- Data Preservation: Penalizing excessive deviation from the original measured projection.
- Latent Bounds: Restricting movement within a bounded neighborhood of the initial encoded state.
- Angular Smoothing: Cyclically smoothing latent states across projection angles to ensure consistency.
Downstream Evaluation: Corrected projections are reconstructed (using TomoPy gridrec), and surfaces are extracted (marching cubes). Performance is evaluated against the no-motion reference using morphology-level metrics: surface area error, connected-component count error, and density-map RMSE/correlation.

Key Contributions

Task-Aligned Framework: The study demonstrates a method for correcting CT motion artifacts that is explicitly optimized for downstream surface-based morphometrics rather than visual image quality.
Preference Learning for Correction: It introduces a mechanism where expert pairwise preferences regarding artifact risk guide the search through a learned residual correction manifold, avoiding the need for differentiable reconstruction or explicit physical motion models.
Bounded Residual Correction: By operating in a residual space with strict clipping and data-preservation penalties, the method mitigates the risk of "hallucinating" specimen structures while correcting motion-induced errors.
Comprehensive Benchmarking: The work provides a controlled synthetic benchmark with five independent seeds and two real-data case studies, comparing the proposed method against visual denoising baselines and random-latent ablations.

Results

Synthetic Benchmark: The preference-guided correction significantly outperformed both a visual baseline (median-stripe suppression + Gaussian denoising) and a random-latent ablation (same autoencoder but random ranking).
- Relative to motion-corrupted data, the method reduced reconstruction RMSE by 67.9% and density-map RMSE by 72.0%.
- Surface-area error was reduced by 69.9% (from 13.4% to 4.0%), and component-count error (duplicated structures) was reduced by 72.0% (from 95.2 to 26.6 components).
- The visual baseline improved reconstruction RMSE by only 20.8%, and the random-latent ablation achieved a 40.8% reduction in density-map RMSE, indicating that the preference signal is crucial for navigating the correction manifold effectively.
No-Motion Controls: When applied to clean, no-motion data, the method produced minimal perturbations (surface-area change < 0.5%, density-map correlation > 0.99), demonstrating that the bounded residual approach does not arbitrarily rewrite clean data.
Real-Data Case Studies: Applied to two real pigeon CT datasets, the method reduced projection ghosting RMSE by ~66% and reconstruction ghost-ridge RMSE by ~65%. Downstream morphometric analysis showed improved agreement with manual reference masks: Intersection-over-Union (IoU) increased from 0.74/0.78 to 0.89/0.91, and spurious crest islands decreased from 6/5 to 1/1.

Significance and Claims
The paper claims to provide proof-of-concept evidence that expert preference signals can guide bounded projection-domain correction for surface-based biological morphometrics. The authors argue that for quantitative biological imaging, the relevant endpoint is not whether a reconstruction looks "cleaner" to a human, but whether the derived surface measurements are less biased.

The study positions this work as a route toward CT imaging workflows optimized for downstream measurement validity. It suggests that while the current implementation focuses on feather motion, the framework is applicable to other scenarios where the correction objective is difficult to define analytically but can be judged by domain experts or downstream analysis (e.g., beam-hardening correction for segmentation accuracy). The authors remain modest regarding the scope, noting that the real-data component serves as application evidence rather than a complete population-level validation, and that the method relies on a single expert's preference model which may require multi-rater validation in future work. The primary contribution is the demonstration that task-aligned, preference-guided optimization can effectively reduce measurement bias caused by motion artifacts without requiring a complete redesign of existing biological analysis pipelines.

Technical Summary: Preference Guided CT Projection Correction for Surface-Based Feather Morphometrics

More like this