Quantum-Inspired Algorithm Uses Thermodynamic Phase Transitions to Sharpen CT Scans

Medical imaging has long faced a stubborn trade-off: the filters and algorithms that clean noise out of a computed tomography scan often blur the very anatomical details that radiologists depend on, while the techniques that sharpen edges tend to amplify grain and introduce artificial halos around organs and lesions. A research team from Nanyang Institute of Technology in China and VSB-Technical University of Ostrava in the Czech Republic now reports a strikingly original way out of this dilemma. In a study published in Complex & Intelligent Systems, Pei Hu, Zishuo Liu, Xiangyu Gao and Jeng-Shyang Pan describe a Binary Quantum Genetic Algorithm with Diversity-Guided Phase Transitions, or BQGA, which borrows concepts from quantum computing and statistical physics to tune CT image enhancement parameters with unusual precision. The work is open access and has been released early as a citable, peer-reviewed accepted manuscript.

At the heart of the method is a reframing of image enhancement as an optimisation problem in a five-dimensional parameter space. The algorithm encodes five CT enhancement settings as quantum bits positioned on the Bloch sphere: window center, window width, contrast gain, noise suppression and edge retention. These are the knobs a radiologist or image-processing pipeline might adjust to make a scan readable, but finding the best combination for any particular volume is notoriously difficult because the parameters interact in nonlinear ways. Conventional genetic algorithms treat this as a global search, evolving a population of candidate solutions over many generations. The problem, the authors note, is that such algorithms typically rely on static rotation schedules that are insensitive to how the population as a whole is converging, which leads to premature diversity collapse in high-dimensional parameter spaces. Once the population loses its diversity, the search stalls in a mediocre solution and never recovers.

BQGA’s central innovation is a closed-loop control mechanism inspired by thermodynamics. At every generation, the algorithm computes a composite diversity score from four complementary population metrics, and this score is mapped to one of three thermodynamic phases: plasma, liquid or crystal. The mapping uses a hysteresis-buffered threshold rule, meaning the system does not flip erratically between phases when the diversity score hovers near a boundary. The active phase then governs three key search parameters simultaneously: the quantum walk step size, the mutation rate and the coin operator. In the plasma phase, the population is highly diverse and the algorithm explores aggressively, taking large quantum walk steps and mutating frequently. In the crystal phase, the population has converged and the algorithm exploits, refining the best solutions with small steps and gentle mutations. The liquid phase sits in between, balancing exploration and exploitation.

The elegance of this design lies in what it eliminates. Traditional metaheuristics require engineers to hand-craft schedules that dictate how exploration should taper off over time, a process that is part art and part guesswork, and a schedule tuned for one class of images may fail badly on another. By letting the measured diversity of the population itself decide the phase, BQGA regulates the exploration-exploitation balance automatically, without any manual schedule design. The authors describe this as closed-loop regulation: the algorithm continuously senses its own convergence state and responds, much as a thermostat senses temperature and adjusts heating. This self-regulation is what allows the method to remain effective across the abdominal and lesion-bearing CT volumes on which it was tested, which vary widely in contrast, noise level and anatomical content.

A second safeguard addresses a specific and familiar artefact of aggressive image enhancement: halos. When an algorithm sharpens edges too forcefully, bright or dark bands appear along boundaries between tissues, which can mislead both human readers and downstream automated analysis. BQGA incorporates a gradient-aware mutation weight that modulates mutation strength according to the overall edge density of the volume being processed. In regions or volumes where edges are dense, the algorithm restrains its mutations, limiting the introduction of halo artefacts during local refinement. This means the enhancement process is not blind to image content; it adapts its own aggressiveness to the structural richness of the data, a feature the ablation experiments later confirm contributes independently to the algorithm’s performance.

To test the method rigorously, the team evaluated BQGA on abdominal and lesion-bearing CT volumes drawn from two widely used public benchmarks, the CHAOS and LiTS2017 datasets. The comparison set was deliberately broad, spanning both metaheuristic optimisers and learning-based methods. Among the metaheuristics were the Quantum Genetic Algorithm, Quantum Particle Swarm Optimisation, Harris Hawks Optimisation and Marine Predators Algorithm, each a well-established competitor in the optimisation literature. Among the learning-based methods were Zero-DCE, a deep network designed for low-light enhancement, and DPM, a diffusion-based approach. BQGA attained the highest mean values of three standard image fidelity metrics: peak signal-to-noise ratio, structural similarity index and universal image quality index. Crucially, the margin over competitors widened on low-contrast, high-noise volumes, precisely the difficult cases where clinical image quality matters most.

Statistical rigor underpins these claims. The authors applied non-parametric tests, which make no assumptions about the underlying distribution of the results, and found that the gains were statistically significant for most fidelity metrics. This matters because image-quality comparisons can be deceptively close; small mean differences across a test set may vanish under proper statistical scrutiny. The team went further, conducting convergence and stability analyses to show that BQGA not only reaches better solutions but does so reliably, without the erratic runs that plague many stochastic optimisers. They also demonstrated practical utility through a downstream Chan-Vese segmentation task, showing that images enhanced by BQGA support better automated segmentation, a common preprocessing step in computer-aided diagnosis pipelines for organs and tumours.

Perhaps the most telling result comes from the ablation analysis, in which the researchers systematically removed components of their algorithm to see which ones carried the weight. The phase-transition mechanism, they found, accounts for the majority of the performance gain. Both the multi-dimensional diversity signal and the gradient-aware weighting contribute independently, meaning the improvements are not redundant layers of the same trick but genuinely complementary mechanisms. This kind of component-level validation is often missing from algorithmic papers, and it gives the community a clear signal about which ideas are worth carrying forward into other domains. The phase-transition framework, in particular, is conceptually portable: any optimisation problem where premature convergence is the enemy could, in principle, adopt a similar diversity-driven phase controller.

The broader significance of the work lies in its synthesis of three previously separate ideas. Quantum-inspired encoding on the Bloch sphere gives the search a rich, continuous representation of candidate solutions. Quantum walks with adjustable step sizes and coin operators provide a principled way to control how candidates move through the search space. And the thermodynamic phase metaphor supplies an intuitive, self-regulating governance layer that replaces fragile hand-tuned schedules. None of these components requires an actual quantum computer; the algorithm runs on conventional hardware and merely borrows the mathematics of quantum states. That makes it immediately deployable in clinical and research settings where quantum hardware is unavailable, while still capturing some of the expressive power that quantum approaches promise.

For radiology, the implications are concrete. CT enhancement parameters directly affect diagnostic confidence, radiation dose trade-offs and the performance of AI systems trained on medical images. An enhancement method that statistically outperforms both classical optimisers and modern deep-learning approaches, especially on the low-contrast, high-noise scans that are hardest to read, could improve the quality of the images on which both human experts and machine-learning models rely. The authors’ demonstration that enhanced images improve Chan-Vese segmentation suggests a natural integration path into existing clinical pipelines. As hospitals accumulate ever larger imaging datasets and automated analysis becomes routine, algorithms that can adapt themselves to each volume, sensing their own convergence and the image’s own structure, may prove to be the quiet workhorses behind clearer scans. The study, published on 1 September 2026 with support from Chinese research programmes on dam concrete damage identification and Henan provincial key scientific research, marks a compelling step in that direction.

Subject of Research: A quantum-inspired genetic algorithm with diversity-guided phase transitions for optimising CT image enhancement parameters

Article Title: Binary quantum genetic algorithm with diversity-guided phase transitions for CT image enhancement

Article References: Hu, P., Liu, Z., Gao, X., & Pan, J.-S. (2026). Binary quantum genetic algorithm with diversity-guided phase transitions for CT image enhancement. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02485-z

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02485-z

Keywords: CT image enhancement, quantum genetic algorithm, phase transitions, quantum walk, diversity control, metaheuristics, medical imaging, image segmentation, optimisation, Bloch sphere, CHAOS dataset, LiTS2017