💻 computer science

Quantum CT via Dynamic Interval Encoding and Prior-Balanced QUBO Reconstruction

This paper proposes a QUBO-based quantum computed tomography framework that utilizes dynamic interval encoding and prior-balanced optimization to overcome binary-variable constraints, enabling high-fidelity grayscale reconstruction on hybrid quantum-classical hardware.

Original authors: Ao Wang, Yikuang Yuluo, Yujie Liu, Shuangyang Zhong, Yuwen Zhang, Zihao Wang, Fenglin Liu, Andreas Maier, Haijun Yu, Yixing Huang

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ao Wang, Yikuang Yuluo, Yujie Liu, Shuangyang Zhong, Yuwen Zhang, Zihao Wang, Fenglin Liu, Andreas Maier, Haijun Yu, Yixing Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, incredibly complex jigsaw puzzle, but you are only allowed to use a tiny, specific set of tools: a limited number of "on/off" switches (binary variables). This is the challenge the paper tackles: creating clear medical images (CT scans) using a new type of computer technology called Quantum Computing.

Here is the story of how the authors solved this puzzle, explained in simple terms.

The Problem: The "Switch" Budget

In a standard CT scan, the computer tries to figure out how dense different parts of your body are (like bone vs. muscle). These densities are like a smooth gradient of colors, from very dark to very bright.

However, the quantum computers they are using can only understand binary data (0s and 1s). To represent a smooth image, you have to turn those colors into a long string of switches.

The Old Way: Imagine trying to describe a painting using a fixed set of 100 switches for every single pixel. If you want a very detailed picture (high precision), you need thousands of switches. But quantum computers have a strict "budget"—they can only handle a few hundred switches at a time. If you try to use too many, the system crashes or gets confused.
The Dilemma: If you use fewer switches to stay within the budget, the picture looks blocky and blurry (like a low-resolution video game). If you try to use more switches for a better picture, the computer can't handle the math.

The Solution: A "Flashlight" Strategy

The authors came up with a clever trick called Dynamic Interval Encoding. Instead of trying to describe the entire picture with a fixed set of switches all at once, they use a "flashlight" approach.

The Flashlight (Local Intervals): Imagine you are looking at a dark room. Instead of trying to see the whole room at once, you shine a flashlight on just one small spot. You ask the computer, "Is this specific spot light, medium, or dark?"
The Guess and Check: The computer makes a guess for that small spot using its limited switches.
Moving the Flashlight: Once the computer solves that small spot, the authors move the "flashlight" to the next spot, but they don't start from scratch. They use the previous guess as a starting point.
The "Boundary Hit" Rule: This is the smart part.
- If the computer's guess hits the very edge of the flashlight's range (meaning the true value might be outside the current guess), the system says, "Okay, we need to look wider!" and expands the flashlight beam.
- If the guess is safely in the middle, the system says, "Great, we're close!" and shrinks the beam to focus on fine details.

By constantly adjusting the size of the "flashlight" based on where the computer is looking, they can build a high-quality image without ever needing to use more switches than the computer can handle.

The "Balanced Chef" (Prior-Balanced Optimization)

To make sure the image doesn't look weird or noisy, the authors added a second ingredient: a Prior. Think of this as a recipe book that tells the computer, "Real human tissue usually looks smooth, but edges (like bone) should be sharp."

However, mixing the "recipe" (the prior) with the "raw data" (the X-ray measurements) is tricky. If you use too much of the recipe, the image looks fake. If you use too little, it looks noisy.

The authors created a "Balanced Chef" system. Before feeding the data to the quantum computer, this system measures the strength of the data and the strength of the recipe, then adjusts them so they work together perfectly. This prevents one from overpowering the other.

The Results: A Clearer Picture

The team tested this method on simulated CT scans with very few X-ray angles (like taking a photo with only a few snapshots instead of a full rotation).

The Competition: They compared their method against standard computer algorithms and other "learning" methods.
The Winner: Their method produced images that were much clearer, with fewer "streaks" and artifacts. It recovered the shapes and shades of the "phantom" (test images) much better than the others.
Real Hardware: They didn't just simulate this; they actually ran it on a real D-Wave quantum computer (a hybrid system that uses both classical and quantum processors). It worked, proving that this "flashlight" strategy is compatible with real quantum hardware.

Summary

The paper claims that by using a dynamic, moving "flashlight" to focus the computer's limited attention on small areas one by one, and by balancing the math between the raw data and the rules of how images should look, they can create high-quality medical images using quantum computers. This solves the problem of quantum computers having a "budget" of switches that is too small for detailed pictures.

What they did NOT claim:

They did not claim this is ready for immediate use in hospitals on real patients.
They did not claim it works on large, noisy, real-world hospital data yet (they used simulated data and small test images).
They did not claim it is faster than current standard methods (the focus was on image quality and feasibility on quantum hardware).

Technical Summary: Quantum CT via Dynamic Interval Encoding and Prior-Balanced QUBO Reconstruction

Problem Statement
Quadratic Unconstrained Binary Optimization (QUBO)-based quantum computed tomography (CT) reformulates image reconstruction as a binary quadratic problem solvable by quantum annealing or hybrid quantum-classical solvers. However, a critical bottleneck exists for grayscale CT: the "binary-variable budget."

Fixed Global Encoding: Traditional approaches assign a fixed global bit-plane encoding to every pixel. As gray-level precision (bit depth) increases, the number of binary variables, quadratic couplings, and coefficient ranges grow rapidly, making the QUBO instance too large for current hardware.
Low-Bit Encoding: Conversely, using low-bit global encodings introduces substantial quantization errors, failing to capture the continuous attenuation coefficients and subtle structural transitions inherent in grayscale images.
The paper identifies the need to balance gray-level expressivity with the constraints of the binary-variable budget available in a single QUBO optimization round.

Methodology
The authors propose a reconstruction framework that replaces fixed global encoding with Dynamic Interval Encoding combined with Prior-Balanced Optimization. The process operates in iterative refinement rounds:

Dynamic Interval Encoding:
- Instead of a global codebook, each active pixel is encoded using a local candidate set centered around its current estimate ( $c_p^{(t)}$ ).
- Within a local search interval $[c_p^{(t)} - r_p^{(t)}, c_p^{(t)} + r_p^{(t)}]$ , the pixel value is represented by a low-bit (e.g., 2-bit) binary vector, yielding four discrete candidate values.
- This allows the system to use a compact number of variables per round while maintaining high effective resolution through iterative updates.
Boundary-Hit-Guided Interval Update:
- After solving the binary optimization, the interval center is updated to the decoded pixel value.
- The interval width ( $r_p$ $r_{p}$ ) is adaptively adjusted based on whether the solution hits the boundary of the current local interval:
  - Boundary Hit: If the solution is at the edge of the interval, it suggests insufficient coverage. The interval is reset to the initial search width ( $r_{init}$ ) to expand the search.
  - Interior Hit: If the solution is within the bounds, the interval is contracted (shrunk) to refine the local precision.
Balanced QUBO Construction:
- The objective function combines a data-fidelity term (projection consistency) and an edge-preserving quadratic prior.
- To ensure stability, the method employs percentile-based normalization to balance the coefficient scales of the data and prior terms before forming the final QUBO matrix. This prevents extreme coefficients from dominating the optimization.
- The prior uses adaptive edge weights derived from a reference image to preserve structural transitions without over-penalizing them.

Key Contributions

Dynamic Interval Encoding: The authors introduce a mechanism to improve effective gray-level expressivity without increasing the per-round bit depth. By updating local intervals across rounds, the method avoids the scalability issues of fixed global high-bit encodings.
Adaptive Update Mechanism: A boundary-hit-guided rule is developed to dynamically switch between search expansion and local refinement, ensuring the algorithm does not get trapped in local minima or fail to cover the true solution space.
Prior-Balanced Optimization: The framework integrates projection-domain data consistency with an edge-preserving prior, utilizing a normalization strategy to stabilize the joint optimization of these competing terms.
Hardware Validation: The formulation is validated on the D-Wave hybrid binary quadratic model (BQM) solver, demonstrating executability on a hardware-backed hybrid quantum-classical backend.

Experimental Results
The method was evaluated on sparse-view (9 and 18 views) and limited-angle fan-beam CT tasks using 32×32 images (and a 72×72 phantom for scalability testing).

Performance: The proposed method (both Simulated Annealing and D-Wave Hybrid variants) significantly outperformed analytic (FBP), iterative (SART, CGN), variational (SART-TV, TV-PDHG), and representation-based baselines (Gaussian, Spener).
- In sparse-view experiments, the Hybrid variant achieved an average PSNR of 40.63 dB and NRMSE of 0.0097, compared to 33.55 dB and 0.0217 for the best conventional baseline (TV-PDHG).
- The method recovered structures and gray-level distributions more faithfully, reducing streak artifacts and residual errors.
Expressivity Analysis: Comparisons showed that dynamic 2-bit encoding outperformed fixed global encodings ranging from 2 to 5 bits. The dynamic approach achieved a PSNR of 51.88 dB in a representative test, whereas fixed 5-bit encoding only reached 30.72 dB, highlighting the efficiency of the local interval strategy.
Ablation Studies: Removing the edge-preserving prior or the term balancing mechanism led to significant degradation in structural continuity and reconstruction quality, confirming the necessity of both components.
Scalability: The method successfully reconstructed a 72×72 phantom (4,460 binary variables) using both local and hybrid solvers, though optimization difficulty increased with image size.

Significance and Claims
The paper claims that the proposed framework addresses the fundamental representation bottleneck in grayscale QUBO CT reconstruction. By decoupling the global bit-depth requirement from the per-round variable budget through dynamic local encoding, the method enables high-fidelity grayscale reconstruction on current quantum and hybrid hardware.

The authors emphasize that the performance gains stem primarily from the effective gray-level representation and stable data-fidelity–prior coupling, rather than specific solver behaviors.
The work serves as a proof of concept, demonstrating that representation design is central to quantum-compatible CT reconstruction.
The paper modestly notes that while the method works on larger instances (72×72), large-scale QUBO CT reconstruction remains a challenge due to binary-variable and bit-depth constraints, identifying this as a key direction for future work involving noisy and measured data.