{"id":"7d09918e-9341-4271-b899-6d92a926d0dd","arxiv_id":"1908.09529","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A two-step method that linearly extends CT sinograms and then uses a U-Net to remove reconstruction artifacts can widen the CT field of view, though the evidence is qualitative and mixed.","lead":"This paper combines linear extrapolation of CT detector channels with a U-Net that cleans artifacts, to widen the usable field of view. The method is tested on clinical scans and compared against an existing Siemens reconstruction method, with mixed qualitative results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training pipeline in Section 2 lacks an explicit truncation step; with standard SFoV targets the EFoV loss forces zero output, contradicting the reported EFoV anatomy.","rationale":"The paper proposes a two-step EFoV method and reports qualitative superiority over HDFoV in some cases. The reader identified simulation-to-real transfer as the weakest assumption. I found a more basic issue: the training target y, as described in Section 2, may not contain any signal in the EFoV region. The loss is restricted to the EFoV region, so if y is a standard SFoV slice (zero outside SFoV), the network is trained to output zero there; the final replacement of the SFoV region by X yields no extension. This is not merely a validation gap; it challenges the validity of the supervised training. The paper never mentions cropping or truncating the sinogram to simulate a smaller FOV, nor does it state that y are wide-FOV reconstructions. This omission makes the method non-reproducible and the central claim unverifiable from the text. The concrete check (train with and without explicit truncation) would determine whether the concern lands. If the no-truncation variant outputs zero in EFoV, the method as described cannot explain the reported results. The paper has other weaknesses (no quantitative metrics, possible selection bias), but the training-data inconsistency is the most load-bearing.","tokens_in":2697,"tokens_out":7899,"duration_ms":80690,"concrete_test":"Implement the training pipeline exactly as described using a public CT dataset: take a reconstructed volume, project to sinograms, apply outer-channel linear extrapolation without any sinogram truncation, reconstruct inputs X, and train a U-Net with the original volume as target, computing the loss only in the EFoV region. Then run the trained network on a test patient and compute the mean absolute value of the output in the EFoV region. If it is near zero (as predicted), the described procedure cannot produce the claimed EFoV anatomy. Repeat with an explicit truncation of the sinogram to a smaller FOV before extrapolation; the second variant should yield non-zero EFoV estimates, confirming that a required step is missing from Section 2.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is that the U-Net learns to synthesize plausible anatomy in the EFoV region. That learning requires training target images y that contain non-zero anatomy outside the simulated SFoV. As written in Section 2, y are 'clinical CT scan slices' projected to sinograms; those sinograms are then extended by linear extrapolation of the outer channels and backprojected to produce inputs X. No step is described that crops or truncates the sinograms to a smaller FOV before extrapolation, nor is it stated that y are reconstructed with a larger FOV than the simulated SFoV. If y is a standard SFoV reconstruction, then its values in the EFoV region are zero. Since the loss is computed only in the EFoV region (Section 2, last paragraph), the optimal network output there is zero. The SFoV region of the network output is replaced by X, so the final image would have no FOV extension at all. This contradicts the qualitative results in Section 3 showing, e.g., an arm in case 3. The method description therefore omits or mis-specifies the essential training-data construction; the reported evaluation cannot be reproduced or interpreted without resolving this ambiguity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-step method for extending the CT field of view. First, the sinogram channels are linearly extrapolated outward, and an EFoV image is reconstructed with FBP, producing characteristic artifacts outside the SFoV. Second, a U-Net is trained to remove these artifacts in image space; the SFoV region is copied from the input, and the network output is statistically matched to the input distribution. The method is evaluated qualitatively on clinical raw data from 11 patients and 4 phantoms, comparing the proposed approach to a prior HDFoV method, with the authors reporting favorable results in three of five illustrated cases. The central claim is that linear extrapolation plus deep-learning artifact removal is an efficient and potentially superior EFoV reconstruction strategy.","tokens_in":2895,"tokens_out":3240,"duration_ms":32689,"significance":"If the central claim holds, the method offers a computationally inexpensive alternative to iterative EFoV reconstruction, which is clinically relevant for radiotherapy planning and imaging of large patients. The combination of a deterministic sinogram-extension step with a learned artifact-reduction step is a sensible division of labor, and the use of a U-Net in image space keeps the runtime low. The paper is an extended abstract, so the contribution is at the proof-of-concept level. The authors are transparent about the qualitative nature of the evaluation and explicitly acknowledge two cases where the proposed method does not outperform HDFoV. However, the current manuscript lacks the training-data details needed to reproduce the method, and the evaluation does not provide quantitative or statistical support for the superiority claim.","major_comments":[{"comment":"The description of the training data construction is ambiguous in a way that is load-bearing for the entire method. The text states that 'a collection of clinical CT scan slices y are projected to their Radon transform to obtain sinograms' and that the loss is 'only computed in the EFoV region.' If y is a standard SFoV reconstruction, then the EFoV region of y is identically zero, and the optimal network output there is zero, which would make the reported EFoV anatomy (e.g., the arm in case 3) impossible. The authors must clarify whether y is reconstructed with a field of view larger than the simulated SFoV, or whether the sinogram is first truncated/cropped to a smaller SFoV before the linear extrapolation is applied. Without an explicit statement of how nonzero anatomy enters the EFoV part of the training targets, the method description is incomplete and the results cannot be interpreted or reproduced.","section":"Section 2, training data paragraph"},{"comment":"The evaluation is entirely qualitative. The authors report that 11 patient and 4 phantom datasets were reconstructed, yet the only evidence presented is five displayed cases, of which three are judged favorable to the proposed method and two are unfavorable. There are no quantitative error metrics (e.g., RMSE, SSIM, or mean absolute error in the EFoV region), even though the phantom datasets could provide ground truth, and no statistical test of the claim that the proposed method is 'superior' to HDFoV in cases 2, 4, and 5. The conclusion that linear extrapolation plus deep learning 'proved to be an efficient method' and 'shows potential to improve quality' is not supported by the evidence shown. The authors should provide quantitative comparisons and, ideally, a reader study or a scoring protocol that substantiates the superiority claim.","section":"Section 3, results"},{"comment":"The network is trained on simulated artifacts generated by linearly extrapolating Radon transforms of clinical CT slices, but it is applied to real scanner raw data in Section 3. The manuscript does not describe whether the training simulation accounts for the actual noise, beam hardening, and truncation characteristics of the scanner, nor does it report the parameters used for the simulated extrapolation (e.g., the extrapolation slope, the amount of padding, or the EFoV/SFoV ratio). If the simulated artifact structure differs from the real artifacts, the U-Net may not remove artifacts on actual scans, and the reported clinical results would not be reproducible. This is a correctness risk that should be addressed by specifying the simulation settings and by showing, at minimum, that the artifact statistics in training match those in the test reconstructions.","section":"Section 2, domain transfer"}],"minor_comments":[{"comment":"The loss is defined as a weighted sum of DSSIM and MSE, but the weights are not reported. Since the relative weighting is a free parameter that can materially affect the trade-off between structural fidelity and pixel accuracy, it should be stated explicitly.","section":"Section 2, loss function"},{"comment":"The order of operations between replacing the SFoV region with X and the statistical adaptation (mean/variance matching) is not made clear. The sentence 'In order to improve the final SFoV/EFoV transition, the network output is adapted to the input distribution' suggests the adaptation is applied before or after the replacement, but the exact sequence should be spelled out.","section":"Section 2, SFoV replacement and normalization"},{"comment":"Figure 2 is difficult to interpret: the caption says 'Data generation pipeline and network architecture,' but the figure appears to mix the U-Net structure with the sinogram-to-image pipeline, and the numbers are not annotated. A separate schematic of the training pipeline and a standard U-Net diagram would improve readability.","section":"Figure 2"},{"comment":"The paper states that volumes were 'checked for consistency and plausibility' but does not describe who performed the checks, how many slices were reviewed, or which criteria were used to judge 'superior' results. A brief description of the evaluation protocol would increase confidence in the qualitative comparison.","section":"Section 3, clinical evaluation protocol"},{"comment":"The abstract and Section 1 mention that the proposed method delivers 'improved results' compared to existing methods, but the introduction cites no quantitative comparison and the body only offers a qualitative comparison to HDFoV. The wording should be aligned with the actual evidence presented.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is an extended abstract from authors affiliated with Siemens Healthcare, and the comparison baseline HDFoV is a Siemens product. This is not inherently problematic, but the qualitative comparison would benefit from an independent evaluation protocol or external validation. The central training-data ambiguity in Section 2 is the main technical blocker; if the authors can clarify that the targets y contain nonzero anatomy outside the SFoV (e.g., by truncating wider-FOV reconstructions), the method is plausible and the revision could be straightforward. The evaluation should be strengthened with at least quantitative phantom results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi X,\n\nThis is a short industrial extended abstract from Siemens on extending the CT field of view. The pairing is genuinely new within the cited literature: linear sinogram extrapolation followed by a U-Net artifact cleanup, wrapped in a ReconCT plugin. The method is simple and cheap, and the authors show five patient-sized cases, explicitly admitting two where their method loses to HDFoV. That honesty is a point in their favor.\n\nWhat the paper does well is bound the scope. They say the EFoV region is non-diagnostic and useful only for radiotherapy planning and large patients. They don't claim a revolution. The architecture itself is a standard U-Net with a slightly unusual replacement of the SFoV region by the input and a mean/variance adaptation; nothing exotic.\n\nThe soft spots are real. The evaluation is entirely qualitative. No error bars, no distance or HU metrics, no ground truth for patients, and the displayed cases are hand-picked. Given that the method loses on two of five shown cases, the claim of 'superior' is not supported. This is an extended abstract, so some absence is expected, but even a table would help.\n\nThe larger issue is the training pipeline as written. The paper says y are clinical CT slices, they are projected to sinograms, the sinograms are extrapolated outward to zero, and reconstructed to give X. The loss is computed only in the EFoV region. If y is a standard SFoV slice, the EFoV region in y is zero, and the optimal output of the network there is zero. That would mean no FOV extension at all, yet the results show an arm reconstructed in case 3. There is an unstated step or a different meaning of y (perhaps an HDFoV reconstruction used as target, or a truncated sinogram that is then extended). Without that step the method is not reproducible and the central claim is internally inconsistent. I think the stress-test note is right and the paper does not answer it.\n\nCitation pattern is fine; the prior EFoV methods are all cited. No code or data, which is typical for industry, but it limits independent verification.\n\nBottom line: the idea is plausible and worth a serious referee despite the problems. The essay needs a proper quantitative evaluation and a corrected/clearer statement of how y and the training targets are constructed. If those are supplied, this could be a useful incremental paper.\n\nI'd send it to peer review, but with referees who will push on the training-data generation.","headline":"Plausible industrial EFoV trick with an honest qualitative report, but the training description has a hole that needs fixing before the results can be believed.","tokens_in":3424,"tokens_out":3601,"would_cite":false,"duration_ms":35204,"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":"CT scans can widen their field of view by stretching sinogram edges and cleaning the artifacts with a U-Net.","keywords":["Extended field of view","computed tomography","sinogram extrapolation","U-Net","artifact reduction","filtered backprojection","truncation artifacts","clinical CT"],"falsifier":"Acquire raw data from a body phantom wider than the scanner field of view, reconstruct it with the proposed pipeline and with HDFoV, and compare both against a ground-truth reconstruction of the same phantom made with a scanner that has a larger field of view; if the U-Net output is not closer to that ground truth than the plain linear extrapolation reconstruction in the extended band, the simulated-to-real transfer fails and the central claim does not hold.","tokens_in":2477,"feed_emoji":"🩻","tokens_out":8801,"duration_ms":80448,"temperature":0.7,"pith_summary":"The paper proposes a two-step way to widen a CT scanner's usable field of view without new hardware. First, the raw detector measurements (the sinogram) are extended by linearly extrapolating the outermost channels toward zero, and a normal filtered backprojection produces an image that is correct inside the original field of view but carries artifacts in the extended region. Second, a U-Net trained on pairs of synthetic extended images and their artifact-free originals removes most of those artifacts, after which the network output is re-normalized to match the input statistics and the original high-quality inner region is copied back in. The authors test this on scanner raw data from 11 patients and 4 phantoms and compare with the existing HDFoV extended-field-of-view method, reporting better-looking results in three of the five shown cases and inferior results in two. The point of the work is that artifact reduction, rather than image inpainting, is enough to produce a clinically usable extended field of view at low computational cost.","feed_headline":"Stretched sinogram edges plus U-Net cleanup widen CT field of view","feed_subtitle":"A low-cost way to image patients wider than the scanner's normal field of view.","key_machinery":"The machinery is a two-stage pipeline built around the Radon transform. A clinical slice is projected to a sinogram, the outermost channels are linearly extrapolated toward zero to match the desired EFoV/SFoV ratio, and filtered backprojection reconstructs an extended image whose outer band contains artifacts. A U-Net, a convolutional encoder-decoder with skip connections, is trained to map those artifact-laden extended images back to the original images, with the loss (mean squared error plus structural dissimilarity) evaluated only in the extended region. The SFoV portion of the network output is discarded and replaced by the original reconstruction, and the output's mean and standard deviation are matched to the input to soften the transition.","core_discovery":"The central claim is that linear extrapolation of sinogram channels followed by deep-learning artifact removal in image space is an efficient and competitive way to extend the CT field of view. The paper's conclusion states that linear extrapolation followed by removal of artifacts in the reconstruction using a deep learning network proved to be an efficient method and shows potential to improve the quality of extended-field-of-view reconstructions. The extended region is not claimed to be diagnostic; it is meant to be sufficient for applications such as radiotherapy planning and imaging of obese patients. On the qualitative comparison with the existing HDFoV method, the proposed reconstructions look superior in cases 2, 4, and 5, but inferior in cases 1 and 3, where the method either introduces artificial structures or fails to reconstruct part of the patient anatomy.","pith_inferences":["A natural extension the paper does not test is varying the noise level and beam-hardening model in the simulated training data; a controlled study could map how much realism the artifact generator needs before the network transfers to raw scanner data.","The failure cases suggest the network is weakest on compact high-contrast anatomy (the missed arm in one case) and strongest on smooth soft-tissue extensions; a phantom study varying tissue contrast in the truncated region could confirm that pattern and guide retraining.","The same two-step idea, linear or model-based extension of incomplete measurements followed by an artifact-removal network, could be transferred to other incomplete-data CT problems such as limited-angle or sparse-view scanning, with the artifact generator replaced by the corresponding forward model."],"forward_implications":["If the claim holds, the extrapolated region becomes usable for non-diagnostic clinical tasks such as radiotherapy planning and imaging of obese patients, where the standard scan field is simply too narrow.","Because the method is one linear extrapolation followed by a single network pass, it offers a computationally cheaper alternative to iterative extended-field-of-view reconstructions.","Framing the problem as artifact reduction rather than image inpainting restricts the space of possible outputs, which should make reconstructed anatomy more plausible than direct extrapolation alone.","The method does not dominate HDFoV in every case: on the five displayed comparisons it is judged superior in three and inferior in two, so it should be considered a selective improvement rather than a universal replacement."],"supporting_citations":[{"why":"Supplies the U-Net architecture used to remove the extrapolation artifacts in image space.","marker":"Ronneberger et al., 2015"},{"why":"Defines the HDFoV method used as the baseline comparison for reconstruction quality.","marker":"Hsieh et al., 2004"},{"why":"Provides an iterative extended-field-of-view reconstruction method cited as an existing alternative.","marker":"Kunze et al., 2007"},{"why":"Provides another efficient extended-field-of-view reconstruction technique used as a comparison point.","marker":"Bruder et al., 2008"},{"why":"Supplies SSIM, whose dissimilarity term enters the training loss.","marker":"Wang et al., 2004"}],"fun_headline_variants":["Linear sinogram outfill plus neural cleanup widens CT scans","Extrapolate sinogram, then U-Net removes artifacts for wider CT","Two-step CT trick: outer channel fill plus deep-learning cleanup","Sinogram edge extrapolation plus deep-learning refines EFoV","Efficient CT field widening via linear sinogram extension and DL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result depends on the assumption that a network trained on artifacts created by stretching simulated projection data of clean CT slices will also clean up artifacts in real scanner measurements, which have noise and beam-hardening effects the simulation may not reproduce.","fun_headline_variants_meta":{"raw":{"variants":["Linear sinogram outfill plus neural cleanup widens CT scans","Extrapolate sinogram, then U-Net removes artifacts for wider CT","Two-step CT trick: outer channel fill plus deep-learning cleanup","Sinogram edge extrapolation plus deep-learning refines EFoV","Efficient CT field widening via linear sinogram extension and DL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000734,"raw_usage":{"total_tokens":3208,"prompt_tokens":797,"completion_tokens":2411,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":413,"completion_tokens_details":{"reasoning_tokens":2321}},"tokens_in":413,"tokens_out":2411,"duration_ms":17530,"temperature":1.0,"reasoning_tokens":2321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:07:55.899801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire raw data from a body phantom wider than the scanner field of view, reconstruct it with the proposed pipeline and with HDFoV, and compare both against a ground-truth reconstruction of the same phantom made with a scanner that has a larger field of view; if the U-Net output is not closer to that ground truth than the plain linear extrapolation reconstruction in the extended band, the simulated-to-real transfer fails and the central claim does not hold.","supporting_citations":[],"review_version":1}