High-Fidelity ROI CT Reconstruction with Limited Quantum Resources via Hybrid Classical-Quantum Refinement

This paper proposes a hybrid classical-quantum framework for high-fidelity CT reconstruction that overcomes quantum resource limitations by using classical methods to generate a stable global image and applying quantum optimization exclusively to a region of interest for local refinement, a strategy demonstrated to yield superior accuracy in reduced-angle settings.

The Gist

The Big Problem: A Quantum Computer That Can't Do It All

Imagine you have a brand-new, super-powerful robot (a quantum computer) that is incredibly good at solving very specific, difficult puzzles. However, this robot has a major limitation: it can only hold a small number of puzzle pieces in its hands at once.

In the world of medical imaging (CT scans), creating a picture of the inside of a body is like solving a giant puzzle. If you try to make the robot solve the entire picture of a large object all at once, it gets overwhelmed. The puzzle is too big, and the robot drops the pieces or makes a mess. This is the main problem the paper addresses: Current quantum computers aren't powerful enough to reconstruct a whole large CT scan image on their own.

The Solution: The "Foreman and Specialist" Team

Instead of asking the robot to do the whole job, the authors propose a hybrid team approach. They split the work into two stages:

1. The Foreman (Classical Computer): First, a standard, old-school computer (which is fast and strong but less "smart" at fine details) builds a rough draft of the whole image. It's like a sketch artist quickly drawing the outline of a building. It gets the general shape right, but the windows and doors might look blurry or a bit wrong.
2. The Specialist (Quantum Computer): Then, the team focuses only on the most important part of the picture—the Region of Interest (ROI). This might be a specific spot where a doctor suspects a tumor or a crack in a machine.
   • The team takes the "rough draft" from the Foreman and asks: "What is missing or wrong in this specific small square?"
   • They give this small, specific question to the Quantum Robot. Because the question is now small and focused, the robot can solve it perfectly, adding high-definition details exactly where they are needed.

How It Works: The "Residual" Trick

The paper uses a clever math trick called a residual projection. Think of it like this:

• Imagine you are trying to clean a dirty window.
• Step 1: You wipe the whole window with a rough cloth (the Classical Computer). It removes the big smudges, but some spots are still streaky.
• Step 2: Instead of wiping the whole window again, you look at the difference between the dirty window and your rough wipe. You see exactly where the streaks are left.
• Step 3: You use a tiny, expensive, high-tech eraser (the Quantum Computer) to clean only those specific streaky spots.

By only asking the quantum computer to fix the "leftover errors" in a small area, the team saves the robot's energy and gets a perfect result for that specific spot.

What They Tested

The researchers tested this idea on three different "phantoms" (fake, computer-generated images of objects):

• Small/Medium Objects: For these, the robot could do the whole job alone, or the team approach worked great. Both methods gave clear pictures of the important area.
• Large/Complex Object: This was the hard test. When the object was big and complicated:
  • If they let the robot try to do the whole thing alone, the result was messy and full of "spot-like" errors (like static on an old TV).
  • If they used the Team Approach (Classical computer for the background + Quantum robot for the specific spot), the result was perfect.

The Key Finding

The most surprising discovery was about the "Foreman" (the classical computer).

• You might think the Foreman needs to be perfect. But the paper found that even if the rough draft had some small errors, the Quantum Specialist could still fix the final picture as long as the rough draft was stable and didn't have wild, crazy mistakes.
• Specifically, using a method called SART (a specific type of classical math) to make the rough draft worked better than using FBP (another common method), even though FBP looked slightly "cleaner" on the background. Why? Because SART created a more stable "foundation" for the quantum robot to build on.

The Bottom Line

The paper concludes that we shouldn't try to force quantum computers to replace our current medical imaging systems entirely. Instead, the best use of this new technology is targeted refinement.

Think of it like a high-end photo editor:

• Use a standard editor to fix the lighting and color of the whole photo (Classical).
• Use a super-powerful, expensive AI tool to sharpen just the eyes or the logo (Quantum).

This approach allows us to get high-quality, detailed images of the most important parts of a scan without needing a quantum computer powerful enough to handle the entire image at once. It's about using the right tool for the right part of the job.

Technical Summary

Technical Summary: High-Fidelity ROI CT Reconstruction with Limited Quantum Resources via Hybrid Classical-Quantum Refinement

Problem Statement
Quantum optimization for Computed Tomography (CT) reconstruction, specifically when formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem, faces significant scalability constraints. The number of binary variables required to represent an image grows rapidly with resolution, making direct whole-image quantum reconstruction infeasible for large or structurally complex objects under current hardware limitations. Furthermore, in practical scenarios such as limited-angle or sparse-view CT, classical iterative methods (e.g., SART) and analytical methods (e.g., FBP) often suffer from artifacts and a loss of fine local detail, particularly when the inverse problem is ill-posed. The challenge lies in achieving high-fidelity reconstruction in specific Regions of Interest (ROI)—such as suspected lesions or defects—without requiring the entire image to be processed by resource-intensive quantum solvers.

Methodology
The authors propose a hybrid Region-of-Interest (ROI) refinement framework designed to operate within limited quantum resources. The pipeline consists of two distinct stages:

1. Coarse Global Reconstruction: A global estimate of the image is first generated using one of three methods: Quantum Tomographic Reconstruction (QTR), Quantum Compressed Sensing Tomographic Reconstruction (QCSTR), Filtered Backprojection (FBP), or the Simultaneous Algebraic Reconstruction Technique (SART).

   • If QTR/QCSTR is used for this stage, it is performed on a reduced-resolution grid to fit within qubit budgets, then upsampled and smoothed.
   • If FBP or SART is used, the reconstruction is performed directly at the target resolution.
   • This stage produces a coarse global estimate ($I_{global}$) intended to capture the background structure rather than serve as the final high-fidelity solution.

2. Residual Projection and ROI Refinement:

   • The ROI pixels in the coarse estimate are zeroed out to create a background-only image ($I_{global \setminus ROI}$).
   • A residual projection image ($P_{res}$) is calculated by subtracting the projection of this background image from the measured experimental projection data ($P$).
   • A second-stage QUBO optimization is formulated exclusively for the ROI. The ROI pixels are encoded using a difference-MAC (Mass Attenuation Coefficient) based discrete representation, where pixel values are expressed as sums of binary variables.
   • The objective is to minimize the squared difference between the residual projection image and the projection generated by the binary-encoded ROI variables.
   • This QUBO problem is solved using a D-Wave hybrid solver, and the optimized ROI is inserted back into the coarse global estimate to produce the final reconstruction.

Key Contributions

• Hybrid Architecture: The paper introduces a strategy that reserves quantum optimization strictly for local refinement, utilizing classical methods to stabilize the global background. This approach decouples the global modeling task from the local high-fidelity recovery task.
• Residual Formulation: By formulating the quantum problem around the residual projection image (the discrepancy between measured data and the coarse background), the method reduces the effective size of the QUBO problem and focuses quantum resources on the most critical structural details.
• Comparative Analysis: The study systematically compares fully quantum pipelines (QTR+QTR) against hybrid pipelines (FBP+QTR and SART+QTR) across varying image sizes and structural complexities under reduced-angle (60-view) settings.

Experimental Results
Experiments were conducted on three synthetic phantom samples: two at 120×120 resolution and one at 180×180 resolution, all with a 60×60 ROI.

• Moderate-Size Samples (120×120): Both the fully quantum pipeline (QTR+QTR) and the hybrid SART+QTR pipeline achieved accurate ROI reconstruction with low Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).
• Complex Sample (180×180): As structural complexity and image size increased, the fully quantum pipeline (QTR+QTR) struggled, producing spot-like artifacts and failing to recover fine details (ROI RMSE: 0.5367). In contrast, the hybrid SART+QTR pipeline, which utilized a classical SART reconstruction for the coarse stage (even using denser 180-view sampling for the coarse stage in this specific case), achieved near-perfect ROI reconstruction (ROI RMSE: 0.0000).
• Coarse Stage Quality: Interestingly, while FBP coarse images often exhibited lower non-ROI error metrics than SART, the SART+QTR pipeline consistently outperformed FBP+QTR in final ROI accuracy. The authors attribute this to the structural stability and "signed" nature of the residual projection; SART provided a more coherent background estimate that minimized distortions in the residual signal used for the QUBO optimization, despite having slightly higher average non-ROI error.

Significance and Claims
The paper claims that the practical advantage of quantum-assisted CT reconstruction does not lie in replacing the entire reconstruction pipeline with quantum methods, but in identifying specific stages where quantum optimization is most effective. The authors posit that a task-separated hybrid strategy is superior: classical reconstruction should be used to provide a reliable, stable global background, while quantum optimization is reserved for targeted local refinement.

The study suggests that for larger and more complex images, the critical factor is not necessarily reducing projection views at every stage, but rather securing a stable global background estimate before applying quantum refinement. This approach allows for high-fidelity localized reconstruction even when the global image is reconstructed under severe data limitations, offering a feasible pathway for precision-focused imaging workflows where only specific diagnostic regions require ultra-high fidelity. The work serves as a proof of feasibility for localized high-fidelity reconstruction under current quantum hardware constraints.