{"id":"0e07cb3f-d4a0-4df0-ad11-2b03c26e18b3","arxiv_id":"2502.04830","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A QUBO-based reconstruction that excludes error-affected sinogram rows achieves low-error reconstructions on small test images, but the paper's 'quantum supremacy' claim is not supported by the comparisons.","lead":"This paper proposes a quantum optimization algorithm for tomographic image reconstruction that masks out corrupted sinogram rows and solves a QUBO model, claiming artifact-free reconstructions from 50 percent of projection angles and with up to 50 percent injected errors. The result is presented as evidence of 'quantum supremacy', but the experiments compare against weak classical baselines and do not establish a quantum speedup.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim of quantum supremacy rests on a comparison against FFT, Gurobi presolve, and untuned SA; the missing control is a competent classical reconstruction from the same error-masked sinogram.","rationale":"The reader's verdict is REJECT, and my concern supports that rejection without changing it. I partially agree with the reader: the rationale correctly notes that FFT is not an optimization baseline, Gurobi is only run to presolve, and SA is not tuned. However, the reader's stated weakest assumption is about error-region detection and solver suboptimality, whereas I see the most load-bearing gap as the missing classical control on the same preprocessed problem. Even if the error detection and solver behaved exactly as reported, the central claim 'these comparative results strongly demonstrate quantum supremacy' would still fail unless the comparison set includes a strong classical reconstruction method that is given the same error-masked or row-deleted sinogram. This is not an internal inconsistency in the QUBO derivation; it is a correctness risk in the inference from the reported tables to the headline claim. A concrete classical baseline test is therefore the decisive check: if TV/DART solves the same masked problem, the quantum solver is not necessary for the demonstrated reconstruction quality, and the word 'supremacy' is unjustified. I keep the verdict unchanged because the reader already rejected the paper, and this concern strengthens that rejection rather than altering it.","tokens_in":15938,"tokens_out":4736,"duration_ms":54907,"concrete_test":"Reproduce the Body 50x50 ring-artifact rows of Table I with the same error-injection and error-masking pipeline, then reconstruct from the masked sinogram using a classical regularized iterative method, e.g., TV-regularized least squares via ADMM or DART, and compute MAE. If any such baseline also achieves MAE 0 (or near 0) on the 30%, 40%, and 50% error cases, the claimed 'quantum supremacy' is not established. As a secondary check, run Gurobi with full branch-and-cut (not presolve only) on the same QUBO for the 10% and 50% ring-artifact cases and record time to proven optimum.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central inference—that MAE 0 on the QUBO solved by the D-Wave hybrid solver demonstrates quantum supremacy—is not supported because no competent classical baseline is run on the same preprocessed problem. After the error-exclusion step in Section III.C masks horizontal error rows, the task becomes a limited-angle/limited-data discrete tomography problem from a sinogram with 50–90% of rows retained; this is exactly the regime where classical TV-regularized least squares or discrete algebraic reconstruction (DART) methods are strong. The supplied baselines are weaker in a way that biases the comparison: FFT/iradon (Section III.B) is a linear filter, not an optimizer, and for ring-artifact cases it is not applied to the error-masked sinogram; Gurobi is run only to its presolve stage (Section III.B states 'we focused on the solution obtained from the presolve stage'); and SA is the untuned D-Wave SimulatedAnnealingSampler. Table I therefore shows only that a commercial hybrid solver beats these weak baselines, not that it outperforms classical reconstruction. The paper's own Section V admits solver suboptimality on the 100x100 tooth case and at 20% error for the head image, so the exact-MAE-0 result is conditional on small instances where the classical comparison was deliberately truncated. The load-bearing missing control is a standard classical discrete-tomography solver applied to the same masked sinograms.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a QUBO-based formulation for parallel-beam tomographic image reconstruction, solves the resulting binary optimization problems with D-Wave's hybrid solver, and compares against FFT/iradon, Gurobi (presolve only), and D-Wave's simulated-annealing sampler. The central claim is that the quantum/hybrid approach achieves 'quantum supremacy' in tomographic imaging by reconstructing images with MAE 0 from sinograms with up to 50% injected ring-artifact errors and from only 50% of the projection angles. Experiments are reported on Shepp-Logan, tooth, body, and head datasets of sizes 30x30 to 100x100, with pixel value ranges from binary to 0-10.","tokens_in":16181,"tokens_out":5809,"duration_ms":64449,"significance":"If the central claim were supported, exact tomographic reconstruction from half the projection angles and from heavily corrupted sinograms would be a practically interesting result for discrete tomography. The QUBO least-squares expansion in Section II.A is standard and the authors provide a public code repository, which are positive elements. However, the claimed quantum supremacy is not established: the classical baselines are truncated or unrepresentative, the error-detection step is unspecified and likely uses knowledge of the injected artifacts, and the paper itself reports solver suboptimality in key cases. The significance of the work as presented is therefore mostly a proof-of-concept that a commercial hybrid QUBO solver can fit small noiseless masked tomography problems, not a demonstration of quantum advantage.","major_comments":[{"comment":"The Gurobi baseline is not a real classical optimization baseline: the authors state that they 'focused on the solution obtained from the presolve stage' rather than allowing Gurobi to solve the QUBO. Presolve is a preprocessing step, so the Gurobi MAE values in Table I do not represent the capability of a state-of-the-art classical QUBO/MILP solver. A fair comparison requires a complete Gurobi run (or an equivalently strong classical optimizer) on the same instances; without this, the hybrid solver's success cannot be attributed to quantum advantage.","section":"Section III.B and Table I"},{"comment":"The error-region detection procedure is not specified. The text says errors are detected by examining density variations between adjacent pixels and locating positions 'where pixel values differ markedly,' but no threshold, window size, or algorithmic rule is given, and the red regions in Fig. 2a appear to be selected from knowledge of the injected errors. Because the QUBO is constructed on the masked domain D\\E, the entire reconstruction depends on this unspecified step. The method is therefore not reproducible, and the MAE-0 results may be an artifact of excluding exactly the corrupted rows.","section":"Section III.C, Eq. (17)"},{"comment":"The missing control is a competent classical discrete-tomography reconstruction applied to the same masked sinogram. After error exclusion, the task is a noiseless or low-noise integer-valued limited-angle reconstruction, a regime where classical methods such as total-variation-regularized least squares or discrete algebraic reconstruction (DART) are known to be strong. FFT/iradon is a linear filtered-backprojection method, not an optimizer, so its poor performance on masked or limited-angle data is expected and does not demonstrate quantum superiority. A concrete test would be to run a TV-regularized or DART reconstruction on D\\E and report the MAE; this is necessary before the Section V claim that the results 'strongly demonstrate the quantum supremacy of our algorithm.'","section":"Section IV.B and Section V"},{"comment":"The paper's own results contradict the claimed robustness. The authors report that the hybrid solver did not reach the global minimum for the 100x100 tooth case (target -5,393,765.588 vs found -5,373,756.428) and failed at 20% error for the head image (target -4,607,247.045 vs found -4,607,144.207). Thus the MAE-0 entries in Table I are instance-dependent and hold only for specific small images and error levels; they do not support a general claim of accurate reconstruction under 'up to 50% error.'","section":"Section IV.A.1 and Section V"},{"comment":"An MAE of 0 on the masked problem is not, by itself, evidence of quantum supremacy. The QUBO in Eq. (17) is a least-squares fit to the remaining sinogram pixels; since the remaining data are noiseless and the pixel values are small integers, any optimizer that reaches the global minimum of the masked QUBO will recover the original image exactly if the remaining projections determine it. These experiments therefore measure whether the hybrid solver can minimize a specific small QUBO, not whether the reconstruction task is classically intractable. The paper provides no complexity argument or scaling analysis showing that the masked QUBO instances are hard for classical algorithms.","section":"Table I and Eq. (17)"}],"minor_comments":[{"comment":"The second sum in Eq. (17) appears to be missing the square on P(theta,s); compare with Eq. (10), where the corresponding term is written as {P(theta,s)}^2.","section":"Eq. (17)"},{"comment":"The parameters of the injected errors are under-specified: the number of corrupted rows, the scaling-factor ranges, and the random seed (if any) are not reported, which hinders reproduction of the ring-artifact experiments in Table I.","section":"Section III.A"},{"comment":"The statement that '50% of the original sinogram data was removed' is not clearly connected to the error-rate definition used elsewhere in the paper; the authors should clarify whether the 50% figure refers to removed projection angles, masked rows, or total sinogram pixels.","section":"Section IV.A.1"},{"comment":"With k ranging from 0 to m, the representable integer range is 0 to 2^{m+1}-1, so the phrase 'less than 2^{m+1}' is slightly off; it should be 'at most 2^{m+1}-1.'","section":"Section II.A, Eq. (4)"},{"comment":"Reference [16] contains the typo 'Hubo formulations' instead of 'QUBO formulations,' and reference [18] lists only K. Jun as author although the text attributes the work to 'Jun and Lee.'","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is best described as a small engineering study of a commercial hybrid QUBO solver applied to discrete tomography, not a demonstration of quantum supremacy. The central claim is unsupported by the experimental design, and the missing classical controls and unspecified error-detection step are load-bearing issues rather than presentation problems. I recommend rejection; a future submission that removes the supremacy claim, specifies the detection algorithm, and compares against a proper classical discrete-tomography baseline could be a modest contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for what it is: an incremental but honest extension of the authors' own QUBO tomography framework, wrapped in a claim of quantum supremacy that the experiments do not support. The genuinely new piece is excluding error-affected sinogram rows from the QUBO domain and showing the approach on ring-artifact and limited-angle cases. The method section is clear, the QUBO algebra is standard least squares, and the code is on GitHub. That is real, reproducible work, and the 25-projection Shepp-Logan result with the hybrid solver matching the original is a legitimate data point.\n\nThe soft spot is the comparison. The central inference—that MAE 0 from the hybrid solver beats classical methods—rests on baselines that are not credible for the same preprocessing. FFT is a linear filter, not a solver for the masked problem; Gurobi is run only to presolve; SA is the untuned D-Wave sampler. No competent classical discrete-tomography or TV-regularized reconstruction is run on the error-masked sinogram. The 50% error case becomes a limited-data problem from the unmasked rows, exactly where classical methods are strong, so the table shows only that a commercial hybrid QUBO solver beats weak baselines. The paper's own Section V admits the hybrid solver misses the global minimum on the 100x100 tooth and at 20% error for the head image, so the exact MAE 0 is conditional and instance-size dependent.\n\nThe error detection threshold is never specified, and the MAE 0 results are fits to the retained sinogram, not predictions against held-out data. That circularity is real but minor if the claim were only 'we can reconstruct from masked sinograms'; it becomes load-bearing because the claim is supremacy.\n\nWhat the paper does well: it states its limitations plainly, describes preprocessing, and ships code, so the empirical part is checkable. Self-citation is not the problem; the prior QUBO work is real and the extension is clearly framed. I would not reject this as nonsense. I would reject the supremacy conclusion and ask for a proper classical baseline on the masked data, plus a specified detection threshold.\n\nWho this is for: people working on QUBO formulations for tomography, and anyone tracking weak-baseline pitfalls in quantum-advantage claims. It deserves a serious referee because the core experiments are reproducible and the flaw is fixable; the right outcome is major revision or rejection, depending on whether the authors will run the missing control.","headline":"An honest, reproducible extension of the authors' own QUBO tomography framework, but the 'quantum supremacy' conclusion is unsupported because the classical baselines are not run on the same masked problem.","tokens_in":16751,"tokens_out":2067,"would_cite":false,"duration_ms":23019,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a QUBO-based quantum annealing algorithm reconstructs tomograms exactly, with mean absolute error zero, even when half the sinogram is corrupted or half the projection angles are missing, while classical FFT and…","keywords":["quantum tomography","QUBO","quantum annealing","tomographic image reconstruction","ring artifacts","limited-angle tomography","missing wedge","quantum supremacy"],"falsifier":"Run the proposed method on a real sinogram from a CT scan with a known detector defect, keeping clinical image sizes, and compare against a defect-free reference; if the density-variation detector fails to isolate the defective rows or the solver returns MAE greater than zero, the central claim fails. A simpler check is to repeat the 50-percent-missing-angle experiment on the 50 × 50 body image with a new random seed and verify MAE 0; the paper already reports the solver missing the global minimum for the 100 × 100 tooth case, so that experiment directly tests the claimed robustness.","tokens_in":15683,"feed_emoji":"⚛️","tokens_out":12017,"duration_ms":116079,"temperature":0.7,"pith_summary":"This paper claims that tomographic image reconstruction can be written as a quadratic unconstrained binary optimization (QUBO) problem and solved by a quantum-annealing hybrid solver so accurately that the reconstructed image matches the original pixel for pixel. In its tabulated experiments the method reports mean absolute error (MAE) 0 for a 50 × 50 body CT image even when half of the sinogram rows carry injected ring-artifact errors and when half of the projection angles are missing, while FFT, a commercial mathematical solver, and simulated annealing produce nonzero errors or no solution on the same cases. The paper calls this quantum supremacy in tomographic imaging. It also reports cases where the solver misses the global minimum, notably a 100 × 100 tooth reconstruction and a 30 × 30 head image at 20 percent error, so the exactness claim is tied to the tested instances. If the claim holds, the approach would make limited-angle and artifact-heavy reconstructions practical in medical imaging and electron tomography.","feed_headline":"Quantum method reports exact CT rebuilds from half the data","feed_subtitle":"QUBO-based annealing handles 50% sinogram corruption or missing angles with mean error zero.","key_machinery":"The load-bearing object is the error-excluded QUBO cost function, Eq. (17): $\\sum_{(\\theta,s)\\in D\\setminus E} \\{(I_P - P)(\\theta,s)\\}^2 - \\sum_{(\\theta,s)\\in D\\setminus E} \\{P(\\theta,s)\\}^2$, where $P$ is the measured sinogram, $I_P$ is the Radon transform of a candidate binary-encoded image, $D$ is the full sinogram domain, and $E$ is the detected error region. Pixel values are expanded as weighted sums of binary qubits, for example $I_{ij} = \\alpha_1 + \\sum_k (\\alpha_k - \\alpha_{k-1}) q_k^{ij}$, so the squared mismatch becomes linear and quadratic in the binary variables. The error region $E$ is found by thresholding the mean absolute density difference between adjacent pixels along the position axis, and the reduced QUBO is handed to a hybrid quantum annealer. The exactness claim is carried by the solver reaching a global minimum whose objective value matches the theoretical minimum $-\\sum_{D\\setminus E} \\{P(\\theta,s)\\}^2$, giving MAE 0.","core_discovery":"The central claim is that excluding corrupted regions of a sinogram before building the QUBO objective lets a quantum annealer find a global minimum that classical optimizers miss. The authors encode each pixel of the target image as qubits whose coefficients follow a mass-attenuation representation, define the cost as the sum of squared differences between the measured sinogram and the Radon transform of the candidate image over the uncorrupted domain $D \\setminus E$, and solve the resulting binary quadratic model with a hybrid quantum-classical solver. Table I reports MAE 0 for all tested ring-artifact rates (10 to 50 percent) and all tested missing-angle rates (10 to 50 percent) on the 50 × 50 body image, and for the 30 × 30 head image at 10 percent error. The same QUBO given to classical tools either returns nonzero MAE or no solution within the time budget. The authors interpret this as demonstrating quantum supremacy for tomographic reconstruction.","pith_inferences":["Beyond the paper, the method suggests a generic recipe for inverse problems: formulate the objective only on the trustworthy subdomain of the data, which could apply to denoising, deconvolution, and sparse-view reconstruction beyond tomography.","Beyond the paper, the quantum-supremacy claim should be read as benchmark superiority on small constructed instances rather than a complexity-theoretic separation, since the paper compares finite solvers on fixed QUBOs.","Beyond the paper, a testable extension is to calibrate the error-row detector on real detector defects, because the paper uses artificial errors and notes that real defects are washed out during resizing."],"forward_implications":["CT scanning could tolerate detector miscalibration without re-scanning, because corrupted projection rows are excluded from the objective rather than corrected.","Limited-angle reconstruction from half the angular range would shorten scan times, reduce radiation dose, and address the missing-wedge problem in electron tomography.","The same error-exclusion recipe should transfer to other inverse problems that can be expressed as QUBO cost functions, such as metal-artifact removal by masking high-density pixels.","The paper's own failure cases show the exactness result is sensitive to problem size and solver behavior, so scaling to clinical images will require larger quantum processors or better solvers."],"supporting_citations":[{"why":"Supplies the original QUBO formulation for CT reconstruction from sinograms that this work extends with error exclusion.","marker":"[17]"},{"why":"Provides the mass-attenuation-coefficient pixel encoding and superposition step used to represent image values with qubits.","marker":"[18]"},{"why":"Shows how systems of linear equations become QUBO problems, grounding the construction of the cost function.","marker":"[15]"},{"why":"Defines the photon mass attenuation coefficients used in the pixel-value representation.","marker":"[24]"},{"why":"Frames the missing-wedge problem in electron tomography that the limited-angle experiments target.","marker":"[23]"},{"why":"Supplies the body CT image dataset used for the MAE 0 ring-artifact and limited-angle experiments.","marker":"[25]"},{"why":"Supplies the head CT image dataset used for the 30 × 30 experiment at 10 percent error.","marker":"[26]"}],"fun_headline_variants":["Quantum annealing rebuilds CT images from half the projections","Exact CT reconstruction from 50% angles via quantum QUBO","Zero-error CT from corrupted sinograms using quantum annealing","Quantum QUBO nails CT rebuilds at half the data","Quantum supremacy: exact CT from 50% missing angles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument relies on two linked assumptions: that corrupted sinogram rows can be reliably detected by abrupt density changes between neighboring pixels, and that after those rows are removed the remaining QUBO is small enough for the quantum hybrid solver to find the true global minimum.","fun_headline_variants_meta":{"raw":{"variants":["Quantum annealing rebuilds CT images from half the projections","Exact CT reconstruction from 50% angles via quantum QUBO","Zero-error CT from corrupted sinograms using quantum annealing","Quantum QUBO nails CT rebuilds at half the data","Quantum supremacy: exact CT from 50% missing angles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1381,"prompt_tokens":931,"completion_tokens":450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":367}},"tokens_in":547,"tokens_out":450,"duration_ms":5045,"temperature":1.0,"reasoning_tokens":367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:17:48.242149+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed method on a real sinogram from a CT scan with a known detector defect, keeping clinical image sizes, and compare against a defect-free reference; if the density-variation detector fails to isolate the defective rows or the solver returns MAE greater than zero, the central claim fails. A simpler check is to repeat the 50-percent-missing-angle experiment on the 50 × 50 body image with a new random seed and verify MAE 0; the paper already reports the solver missing the global minimum for the 100 × 100 tooth case, so that experiment directly tests the claimed robustness.","supporting_citations":[{"cited_title":"A highly accurate quantum optimization algorithm for CT image reconstruction based on sinogram patterns,","cited_arxiv_id":null,"evidence_quote":"Supplies the original QUBO formulation for CT reconstruction from sinograms that this work extends with error exclusion."},{"cited_title":"Quantum optimization algorithms for CT image segmentation from X-ray data","cited_arxiv_id":"2306.05522","evidence_quote":"Provides the mass-attenuation-coefficient pixel encoding and superposition step used to represent image values with qubits."},{"cited_title":"QUBO formulations for a system of linear equations,","cited_arxiv_id":null,"evidence_quote":"Shows how systems of linear equations become QUBO problems, grounding the construction of the cost function."},{"cited_title":"Photon mass attenuation and energy - absorption coefficients,","cited_arxiv_id":null,"evidence_quote":"Defines the photon mass attenuation coefficients used in the pixel-value representation."},{"cited_title":"Electron tomography: a three -dimensional analytic tool for hard and soft materials research,","cited_arxiv_id":null,"evidence_quote":"Frames the missing-wedge problem in electron tomography that the limited-angle experiments target."},{"cited_title":"Radiology Data from The Cancer Genome Atlas Lung Adenocarcinoma [TCGA - LUAD] collection","cited_arxiv_id":null,"evidence_quote":"Supplies the body CT image dataset used for the MAE 0 ring-artifact and limited-angle experiments."},{"cited_title":"Computed Tomography (CT) of the Brain","cited_arxiv_id":null,"evidence_quote":"Supplies the head CT image dataset used for the 30 × 30 experiment at 10 percent error."}],"review_version":1}