{"id":"6ed1995c-acc4-4d37-b5cc-3ef27703e291","arxiv_id":"2504.20654","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A region-wise iterative QUBO refinement method reconstructs 100x100 tomographic phantoms with 2,500 to 7,500 qubits, but the dose-reduction claim is extrapolated from an untested downscaling ratio.","lead":"This paper proposes a quantum-assisted CT reconstruction pipeline that starts from a downscaled sinogram and refines the image region-by-region using a D-Wave hybrid quantum solver. The authors report accurate 100x100 phantom reconstructions with fewer qubits, but the headline claim of 90% radiation dose reduction from a 50x50 start is not supported by the experiments.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The refinement objective (Eqs. 6-9) fits each region to the original full-resolution sinogram P, so the advertised 90% dose reduction is not supported by the method or the experiments.","rationale":"The reader's weakest assumption exactly identifies the load-bearing weakness: the refinement stage uses the original full-resolution sinogram P, so reconstructing from a downscaled version of an already-acquired sinogram cannot reduce radiation dose. My read of Eqs. (6)-(9) confirms this. The 500×500 / 50×50 dose claim is a size-ratio calculation, not a measured outcome, and the only sparse-view experiment reduces views by 50%, not 90%. This concern is central because the abstract and introduction advertise radiation-dose reduction as a primary contribution. The qubit-reduction claim is better supported: the 100×100 binary experiment uses four 50×50 regions with one qubit per pixel (2,500 qubits), and the integer-valued experiment uses three qubits per pixel (7,500 qubits), both within the stated limits. However, the paper provides no code, no baselines, and no quantitative image-quality metrics, only visual similarity and energy differences, which further weakens verification but is secondary to the dose issue. Because the concern aligns with the reader's conditional verdict rather than introducing a new objection, I recommend no change: the paper should remain conditional on removing or rigorously supporting the dose-reduction claim, ideally by a true sparse-acquisition experiment at the claimed 10% dose level.","tokens_in":13664,"tokens_out":4354,"duration_ms":47652,"concrete_test":"Run the proposed region-wise refinement pipeline on a 500×500 binary Shepp-Logan phantom using a sinogram physically acquired at 10% of the full projection budget (e.g., 10 projection angles, or an equivalent sparse/downsampled detector readout), and use only that reduced sinogram in Eqs. (6)-(9) — never the full-resolution P. Report reconstruction error (MSE and SSIM) against the phantom. If the final image requires access to the full-resolution P during refinement, or if the sparse-acquisition reconstruction fails, the 90% dose-reduction claim is not valid. A minimal analytic check is to replace P with the downscaled sinogram in Eq. (6) at every refinement stage and compare convergence; this would directly reveal whether the pipeline depends on full-dose data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central advertised benefit — up to 90% reduction in radiation dose — is not supported by the algorithm as written. In Section 2.2.1, the region-wise refinement objective is defined against the original, full-resolution sinogram P(θ,s): Eq. (6) minimizes the squared difference between the superposed sinogram and P − P_z, and Eq. (9) expands this in terms of the measured projection values P(θ,s). The downscaled sinogram P̃ is used only to produce the initial low-resolution reconstruction; once refinement begins, each region is fit to the full-dose measurement P, not to a low-dose acquisition. Since the radiation dose is determined by how P is physically acquired, downscaling P after acquisition cannot reduce patient exposure. The 90% figure is arithmetic on image sizes, (1 − 50/500), and the claimed 500×500 experiment is not reported anywhere in the manuscript. The only reduced-acquisition experiment uses 50 projection angles instead of 100, a 50% reduction in views, not 90%. Unless projection data are physically acquired at reduced resolution or with sparse angles and the refinement is run against that reduced-acquisition sinogram, the dose-reduction claim is unsupported. This does not invalidate the qubit-reduction result for the demonstrated 100×100 cases, but it removes the second half of the abstract's central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid quantum-classical tomographic reconstruction framework intended to reduce both qubit counts and radiation exposure. The method first downscales a sinogram, solves a low-resolution QUBO using the D-Wave hybrid solver, upscales the resulting image, and then iteratively refines spatial subregions by solving local QUBO problems derived from the original full-resolution sinogram. Two variants are presented: full-view sinogram downscaling and sparse-view acquisition with detector-axis downscaling. Experiments on 100x100 binary and integer-valued Shepp-Logan phantoms show visually accurate reconstructions using 2,500 and 7,500 qubits per region, respectively. The abstract and discussion further claim that reconstructing a 500x500 image from a 50x50 initialization could reduce projection data, and hence radiation dose, by up to 90%.","tokens_in":14011,"tokens_out":6514,"duration_ms":71974,"significance":"If the central qubit-reduction result holds, the contribution is a useful incremental step for quantum-annealing-based tomography: replacing a full 10,000-variable QUBO with four 2,500-variable regional QUBOs is practically meaningful for current D-Wave hardware, and the authors provide detailed D-Wave energy logs and a transparent iterative procedure for two phantom cases. However, the advertised clinical benefit, up to 90% radiation dose reduction, is not supported by the algorithm as written or by the experiments. The 500x500 case is never run, and the refinement stage fits each region to the original full-resolution sinogram rather than to a physically reduced-dose acquisition. The paper therefore contributes a qubit-reduction technique and a data-reduction heuristic, but not a demonstrated low-dose imaging protocol.","major_comments":[{"comment":"The claimed radiation-dose reduction is unsupported because the refinement objective in Eqs. (6)-(9) minimizes the difference between the superposed region sinogram and the original full-resolution sinogram P. Patient dose is determined by how P is physically acquired, so downscaling P after acquisition does not reduce exposure. The sparse-view experiment reduces the number of projection angles from 100 to 50, which is a 50% reduction in views, and detector-axis pooling does not by itself lower dose; no experiment approaches the claimed 90% reduction. Please either revise the abstract and discussion to describe data-volume reduction in post-processing, or redesign the pipeline so refinement is performed against a sinogram physically acquired at reduced resolution and demonstrate the dose saving with an experiment.","section":"Section 2.2.1, Eqs. (6)-(9); Abstract; Section IV"},{"comment":"The 90% reduction figure is arithmetically inconsistent with the paper's own downscaling model. Section 2.2.1 downscales both the angular and detector dimensions, with N_S = n_S d1 and M_S = m_S d2, so the number of sinogram elements scales as (n/N)^2, not as n/N. For N=500 and n=50, two-dimensional downscaling gives a 99% reduction in sinogram elements, while the claimed 90% corresponds only to one-dimensional detector downscaling. The paper should define precisely which acquisition parameters are assumed to scale with dose and reconcile the formula (1 - n/N) with the two-dimensional downscaling used in the method.","section":"Section I and Section IV, 90% formula"},{"comment":"The accuracy claims rely on visual inspection and on the gap between the target minimum energy and the D-Wave minimum energy. The energy gap is a self-consistency check of the QUBO solve, not a measure of tomographic fidelity, because the target minimum is computed from the same objective being solved. No quantitative image error (e.g., MAE, MSE, SSIM) is reported for the final reconstructions. Please add such metrics, and state explicitly whether the binary reconstruction is exactly equal to the phantom or merely visually identical.","section":"Section III, Table I and Figs. 3-4"}],"minor_comments":[{"comment":"The phrase \"moving the constant term to the left-hand side\" is confusing; the constant term is subtracted from the expanded expression to obtain the QUBO form. Please rephrase for clarity.","section":"Section 2.2.1, Eq. (10)"},{"comment":"Target and D-Wave energies are reported with up to ten decimal places, which likely exceeds the precision justified by the sinogram data and the solver. Please report energies with an appropriate number of significant digits.","section":"Table I"},{"comment":"The caption says \"1 × d vertical patch,\" but the text describes pooling along the detector/position axis. Please clarify the orientation and terminology so the figure matches the text.","section":"Figure 2(b)"},{"comment":"The estimates for maximum reconstructible image size under a 10,000-qubit constraint and the O(10^6) annealing computations are stated without derivation. Please either provide a short derivation or label these as rough heuristics with clear assumptions.","section":"Section IV, simple decomposition"},{"comment":"The section titled \"APPENDIX AND THE USE OF SUPPLEMENTAL FILES\" is empty and should be removed, and the header date \"NOVEMBER 2020\" does not match the submission year.","section":"End matter"}],"recommendation":"major_revision","confidential_remarks":"The qubit-reduction contribution is real for the demonstrated 100x100 cases, but the clinical framing in the abstract and discussion substantially overstates what is shown. If the dose-reduction claims are removed or properly qualified, the paper may be suitable as a methods contribution; as it stands, the central advertised benefit is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the qubit-saving idea is real, the dose-saving headline is not. The paper's refinement stage fits each region to the original full-resolution sinogram P (Eqs. 6-9), so downscaling P after acquisition doesn't reduce the radiation a patient receives. The 90% figure is just (1 - 50/500), and no 500x500 experiment is reported.\n\nWhat's actually new: the region-wise iterative refinement pipeline, where a low-res QUBO reconstruction is upscaled and then refined region-by-region with the rest of the image fixed. That's a sensible engineering extension of their earlier QUBO work (refs 6, 10, 11), and the demonstrated 100x100 results are decent: 2,500 qubits for binary, 7,500 for integer-valued, with D-Wave energies landing close to the target minima. The finding that nearest-neighbor upsampling creates artifacts that block convergence, and that smoothing helps, is a useful practical detail.\n\nSoft spots, in order of severity:\n\n1. The dose-reduction claim is unsupported. The refinement objective in Section 2.2.1 is defined against P, the full-resolution measured sinogram. Downscaling after measurement is a computational choice, not an acquisition choice. Unless the low-resolution sinogram is physically acquired (or fewer angles are used) and the refinement runs against that reduced-acquisition data, the 90% dose claim is arithmetic on image sizes. The only reduced-acquisition experiment uses 50 angles instead of 100—a 50% view reduction, not 90%.\n\n2. Missing support: no code, no data, no quantitative image metrics (MSE, SSIM), no classical baselines (e.g., FBP or TV-constrained reconstruction). The figures are visual only, and 'identical to the original' is asserted, not measured.\n\n3. The target minimum energies are computed from the same formulation that is being solved, so the D-Wave agreement mainly shows the solver can find the stated minimum, not that the reconstruction is correct. That's fine as a solver check, but it doesn't validate the imaging.\n\nThe core technical contribution—qubit-efficient region-wise refinement—holds up for the demonstrated 100x100 cases. The paper is for researchers working on quantum annealing for CT reconstruction. It deserves a serious referee, but the authors need to remove or fundamentally reframe the dose claim, add real sparse-acquisition experiments, and bring in standard metrics and baselines before this is publishable.\n\nI'd send it to review with a clear request for those revisions.","headline":"The region-wise QUBO refinement genuinely cuts qubit count for 100x100 reconstructions, but the advertised 90% dose reduction does not follow from the method—the refinement stage uses the full-resolution sinogram.","tokens_in":14507,"tokens_out":2555,"would_cite":true,"duration_ms":23813,"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":"Region-by-region quantum refinement reconstructs a 100×100 CT image from 2,500 qubits.","keywords":["quantum annealing","tomographic reconstruction","QUBO","region-wise refinement","sinogram downscaling","sparse-view CT","low-dose imaging","qubit efficiency"],"falsifier":"Measure the radiation exposure of the actual acquisition that produces the full-resolution sinogram $P$ used in Equations (6)–(9), then check whether the downscaling step changes delivered dose; if the dose is fixed by acquiring $P$, the claimed 90% reduction is absent. A direct alternative is to acquire a true 50×50 sinogram on a phantom and test whether the refinement pipeline still matches the 100×100 reconstruction.","tokens_in":13469,"feed_emoji":"🩻","tokens_out":8885,"duration_ms":82956,"temperature":0.7,"pith_summary":"The paper proposes a way to make quantum-assisted tomographic reconstruction scale to higher resolutions without a full-image quadratic unconstrained binary optimization (QUBO) problem. Its core idea is to reconstruct a small image from a downscaled sinogram, upscale it, and then refine one fixed-size region at a time while the rest of the image stays fixed. On a 100×100 binary Shepp-Logan phantom this recovers the exact target image using four 50×50 region solves, i.e. 2,500 qubits per solve; on a 100×100 integer-valued phantom with pixel values 0–3 it uses 7,500 qubits and closely matches the target. The authors further argue that starting a 500×500 reconstruction from a 50×50 initialization could cut projection data and radiation dose by about 90%.","feed_headline":"Region-wise quantum refinement rebuilds 100×100 scans with 2,500 qubits","feed_subtitle":"A downscaled start plus local QUBO solves may also cut projection data for larger scans.","key_machinery":"The load-bearing mechanism is the region-wise QUBO built from a zero-masked sinogram difference. For a selected region $S_l$, the pixels are encoded with $m$ binary variables each, the region's contribution is isolated as $D(\\theta,s)=P(\\theta,s)-P_z(\\theta,s)$, and the objective $\\sum_{\\theta,s}(I_P(\\theta,s)-D(\\theta,s))^2$ is expanded into $\\mathbf{x}^\\top\\mathbf{Q}\\mathbf{x}+\\mathbf{c}^\\top\\mathbf{x}$. Because only $n^2 m$ variables enter each solve, qubit count depends on region size rather than final image size; repeated passes over all regions, with interpolation and filtering between passes, drive convergence.","core_discovery":"The central discovery is that the qubit bottleneck in quantum tomography can be bypassed by treating reconstruction as coarse-to-fine refinement over spatial regions rather than as one large binary optimization. Each refinement step builds a local QUBO whose objective is the squared difference between the projection of a qubit-encoded region and the target contribution $D(\\theta,s)=P(\\theta,s)-P_z(\\theta,s)$ obtained by subtracting a zero-masked sinogram from the original sinogram. With 100 projection angles, a 100×100 binary phantom is reconstructed from a 50×50 start after two passes over four regions, and a sparse-view 50-angle integer phantom is recovered after smoother interpolation and Gaussian filtering. These results are offered as evidence that 2,500 and 7,500 qubits respectively suffice for 100×100 binary and integer-valued images, and that the pipeline extends naturally to sparse-view acquisition.","pith_inferences":["The dose-reduction claim is not established by the reported experiments: those experiments acquire a full-resolution sinogram and then downscale it, so the radiation exposure is set by the full acquisition, not by the downscaled reconstruction input.","Because the region-wise loop is a block-coordinate descent over spatially disjoint blocks, a classical QUBO solver could run the same decomposition; comparing classical and quantum solvers on the same region subproblems would isolate what quantum annealing actually adds.","A natural testable extension is to acquire a true low-resolution sinogram at the detector level and check whether the same refinement pipeline recovers the 100×100 target; if it does, the dose-reduction claim would move from ratio arithmetic to measurement."],"forward_implications":["A 100×100 binary tomographic image can be reconstructed with 2,500 qubits, and a 100×100 integer-valued image with 7,500 qubits, under the four-region scheme demonstrated.","Sparse-view sinograms with half the projection angles pass through the same region-wise refinement pipeline without algorithmic changes.","If the low-resolution sinogram is acquired directly, the paper's $(1-n/N)$ formula implies starting a 500×500 reconstruction at 50×50 would require roughly 90% less projection data and radiation dose.","Overlapping or adaptively chosen regions are supported, so regions of clinical interest could be refined selectively rather than the whole image.","Initialization smoothness matters: nearest-neighbor upscaling stalls refinement, while smoother interpolation plus Gaussian filtering ($\\sigma=1$) lets the second pass converge."],"supporting_citations":[{"why":"supplies the base QUBO-based quantum CT reconstruction formulation and the radix-2 per-pixel binary encoding used in the region solves.","marker":"[10]"},{"why":"supplies the MAC-aware qubit encoding for known attenuation coefficients and the extension used for the integer-valued phantom.","marker":"[11]"},{"why":"supplies the iterative QUBO refinement strategy with limited qubits that the region-wise loop adapts and extends.","marker":"[6]"},{"why":"provides the hybrid quantum-classical solver service that carries out all QUBO minimizations in the experiments.","marker":"[8]"},{"why":"provides the earlier claim that quantum tomography can reconstruct images from partially ideal projection data, motivating low-data acquisition.","marker":"[24]"}],"fun_headline_variants":["Quantum tomography from 90% less data via coarse-to-fine refinement","Qubit-saving tomography: 2,500 qubits to rebuild 100×100 scans","2500 qubits for 100×100 tomography via region-wise QUBO","Low-dose quantum imaging: region-wise refinement needs fewer qubits","Tomography with 2,500 qubits: region-wise refinement scales to 100×100"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dose-reduction claim rests on the assumption that reducing the sinogram that enters the reconstruction also reduces what is physically acquired; the paper's experiments downscale an already-acquired full-resolution sinogram, so the dose saving is assumed rather than measured.","fun_headline_variants_meta":{"raw":{"variants":["Quantum tomography from 90% less data via coarse-to-fine refinement","Qubit-saving tomography: 2,500 qubits to rebuild 100×100 scans","2500 qubits for 100×100 tomography via region-wise QUBO","Low-dose quantum imaging: region-wise refinement needs fewer qubits","Tomography with 2,500 qubits: region-wise refinement scales to 100×100"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000906,"raw_usage":{"total_tokens":3895,"prompt_tokens":940,"completion_tokens":2955,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2849}},"tokens_in":556,"tokens_out":2955,"duration_ms":19485,"temperature":1.0,"reasoning_tokens":2849,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:23:37.435571+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the radiation exposure of the actual acquisition that produces the full-resolution sinogram $P$ used in Equations (6)–(9), then check whether the downscaling step changes delivered dose; if the dose is fixed by acquiring $P$, the claimed 90% reduction is absent. A direct alternative is to acquire a true 50×50 sinogram on a phantom and test whether the refinement pipeline still matches the 100×100 reconstruction.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the base QUBO-based quantum CT reconstruction formulation and the radix-2 per-pixel binary encoding used in the region solves."}],"review_version":1}