The problem the team set out to solve is one that anyone working with cone-beam CT (CBCT) knows intimately. CBCT, originally developed for dental diagnostics, has become a workhorse of image-guided radiation therapy, interventional radiology and surgical navigation because it delivers high-resolution 3D images directly in the treatment room at lower radiation doses and faster scan times than conventional CT. But the traditional reconstruction method, the Feldkamp-Davis-Kress (FDK) algorithm, demands a large number of X-ray projections. When clinicians cut the number of projections to reduce dose — a critical safety goal — FDK produces noisy, artifact-riddled images.

Iterative reconstruction methods offer a way out. Instead of applying a single filtering step, they repeatedly refine an image estimate: the current guess is forward-projected to simulate what the scanner would measure, the simulated projections are compared with the actual measured data, and the difference is back-projected to update the image. Regularization terms, often based on the total variation (TV) norm, stabilize the solution by encouraging smooth tissue regions while preserving sharp anatomical edges. Algorithms such as ASD-POCS and its adaptive-weighted variant AwPCCS can produce strikingly good images from sparse, low-dose data — but only if their hyperparameters are set just right.

That is the catch. ASD-POCS alone carries eight tunable parameters, including the maximum number of iterations, the number of TV minimization steps, a data inconsistency tolerance, and several step-size and reduction factors spanning value ranges from a thousand candidate settings down to just ten. Even tiny deviations from the optimal combination can visibly degrade image quality. Manual tuning is tedious, time-consuming, and prone to observer variability, and any change in machine, organ, or radiation protocol typically forces clinicians to start over. Prior automated approaches have mostly optimized a single parameter, most often the regularization weight, and nearly all have relied on reference-based objective functions that assume a ground-truth reconstruction exists — an assumption that collapses in real clinical scenarios.

The new framework, called Search Space Aware Crow Search Algorithm (SSA-CSA), attacks both gaps at once. It is built on the crow search algorithm (CSA), a bio-inspired metaheuristic that models how crows hide food and follow flockmates to steal their caches. In the original CSA, each crow’s position encodes a candidate solution; here, each crow carries a vector of hyperparameter values, one per tunable parameter, allowing the entire parameter set of a reconstruction algorithm to be optimized simultaneously. The researchers reformulated all five core components of a metaheuristic — population initialization, exploitation, exploration, the balance between local and global search, and the fitness function — into a unified, reference-free pipeline.

Several technical innovations make the enhanced algorithm work. For initialization, the team developed a Chaotic Diagonal Linear Uniform (CDLU) scheme that combines diagonal stratified sampling with a deterministic sine chaos map, ensuring the initial flock of crows covers the enormous search space evenly without the ordering problems of standard diagonal sampling. For local search, crows now follow members of a dynamically shrinking ‘superior set’ of top performers identified via Pareto dominance, and random step factors are replaced with chaotic values. For global search, the crows gain ‘search space awareness’: a weight map is progressively built up around the best-performing regions of each parameter dimension, and exploration is steered by weighted random sampling rather than blind uniform randomness. The balance between exploitation and exploration is likewise made fitness-aware, so high-quality crows exploit while weaker ones explore, with the threshold tightening as the run progresses.

Equally important is the fitness function that guides the whole search. Because no reference image exists, the team designed a multi-objective, no-reference score combining the Signal-to-Noise Ratio (SNR), which captures noise reduction, with the high-frequency energy ratio (HFER), a Fourier-domain measure of image sharpness. The two metrics inherently conflict — smoother images are less noisy but blurrier — so the framework seeks the best trade-off point via a weighted sum. Ablation experiments on the publicly available SophiaBeads microCT dataset showed that this SNR-plus-HFER fitness outperformed alternatives based on SNR alone, Laplacian variance, or HFER alone when judged by two state-of-the-art learning-based image quality models, CHILL@UK and RPI_AXIS, and correlated strongly with those models’ scores.

The evaluation was unusually thorough. The framework was tested on six datasets from four different imaging systems, including a Nikon Custom Bay CT scanner, a medPhoton ImagingRing m device, and Philips Allura and Azurion C-arms, spanning glass-bead phantoms, line-pair CatPhan tests, a 3D-printed anthropomorphic thorax, a brain phantom with an inserted metal screw, a reconfigurable ChaosCube phantom, and simulated projections from five real patient CT volumes in the LIDC-IDRI lung database. Across ASD-POCS, AwPCSD and the prior-image-based PICCS algorithm, SSA-CSA consistently beat both manual parameter setting and the original CSA, with average improvements of roughly 3.9 percent in fitness, 4.3 percent on CHILL@UK and 3.9 percent on RPI_AXIS. A blind reader study by an expert radiologist with twenty years of CT experience ranked the SSA-CSA reconstructions first and awarded them the highest quality scores, including a perfect 5 out of 5 for the PICCS brain reconstruction from only 20 percent of the full projection set.

Perhaps the most clinically significant finding concerns generalizability. When the optimal parameter set found for one lung CT volume was applied to other patients and to different dose levels — from 100 percent of projections down to a severely undersampled 25 percent — image quality remained consistently good, indicating that a single optimization run can serve an entire class of similar imaging tasks. The same held for six different configurations of the ChaosCube phantom. In benchmark tests against eleven other optimization algorithms across nine standard functions, SSA-CSA showed fast convergence, low run-to-run variability attributable to its deterministic initialization, and statistically significant superiority in most pairwise Wilcoxon comparisons, even under the deliberately constrained budget of 25 individuals and 30 iterations chosen to reflect real clinical hardware limits.

The authors are candid about limitations: like all metaheuristics, SSA-CSA cannot guarantee a global optimum, its fitness weights require some experimental tuning, and its search-space awareness adds computational overhead — though this is negligible next to the cost of the reconstructions themselves and is incurred only once, offline. Future work may replace the hand-crafted fitness with a learned model or swap the metaheuristic for deep reinforcement learning. For now, the framework offers something clinics have lacked: a practical, operator-independent way to squeeze the best possible image quality out of low-dose iterative reconstruction, no reference image required.

Subject of Research: Automatic hyperparameter optimization for iterative cone-beam CT image reconstruction using an enhanced crow search metaheuristic algorithm

Article Title: Reference-free framework for automatic parameter optimization in iterative reconstruction using a novel search-space-aware crow search algorithm

Article References: MohammadiNasab, P., Biguri, A., Steininger, P., Keuschnigg, P., Lamminger, L., Lach, A., Islam, S. M. R. S., Breger, A., Karner, C., Schönlieb, C.-B., Birkfellner, W., & Hatamikia, S. (2026). Reference-free framework for automatic parameter optimization in iterative reconstruction using a novel search-space-aware crow search algorithm. Results in Engineering, 32, Article 113147. https://doi.org/10.1016/j.rineng.2026.113147

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113147

Keywords: cone-beam CT, iterative reconstruction, crow search algorithm, metaheuristics, hyperparameter optimization, low-dose imaging, image quality assessment, total variation regularization, ASD-POCS, PICCS, medical imaging, radiation dose reduction