{"id":"f0d7605c-c6dc-419a-b932-86a6073bfba2","arxiv_id":"2501.14013","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Fusing non-contrast, arterial, and portal venous CT with a 3D progressive fusion network improves portal venous quality and pancreas segmentation by about 3%.","lead":"A deep learning model used three CT scan phases to improve the quality of the portal venous phase and boosted automated pancreas segmentation by about 3% compared with degraded scans. The work is a proof of concept that multiphase information can help downstream organ segmentation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multiphase contribution not validated: 3D-PFNL shows no significant improvement over single-phase 3D-RCAN, so the central claim of multiphase benefit is unsupported.","rationale":"I read the paper as a proof-of-concept that multiphase CT can enhance portal venous quality enough to benefit pancreas segmentation. The most load-bearing condition for that claim is that the multiphase fusion itself contributes measurably over a single-phase enhancement. The paper's own comparison with 3D-RCAN, a single-phase method, shows no significant difference in segmentation (and worse PSNR/SSIM), which directly undercuts the claimed advantage of using multiple phases. This is an internal gap, not an external generalization issue; it can be resolved by computing the missing pairwise statistics or running an ablation. The reader's weakest assumption about synthetic degradation is real but secondary: even if the degradation model were perfect, the paper still fails to show that multiphase information matters. The verdict remains CONDITIONAL because the paper could be salvaged by supplying the missing comparison; however, without it, the central claim is unsubstantiated. I partially agree with the reader because both of us flag the multiphase contribution as not isolated, but I elevate this to the primary concern rather than the synthetic-data fidelity.","tokens_in":7505,"tokens_out":3846,"duration_ms":33893,"concrete_test":"Perform a paired Wilcoxon signed-rank test on per-patient Dice and NSD between 3D-PFNL and 3D-RCAN restorations on the 16 test subjects. If p ≥ 0.05, the multiphase advantage is not statistically established. Additionally, ablate 3D-PFNL by removing the non-contrast and arterial branches (i.e., single PV input) while keeping the same architecture and loss; if the three-phase version does not significantly beat this single-phase ablation, the claimed benefit of multiphase input is not demonstrated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central novelty is that fusing non-contrast and arterial phases improves portal venous quality enhancement. Yet Table I shows 3D-PFNL has lower PSNR/SSIM than single-phase 3D-RCAN (28.16/0.8733 vs 28.77/0.8794). In the segmentation proxy, 3D-PFNL Dice/NSD are only 0.9 and 1.0 percentage points higher than 3D-RCAN, and the Discussion explicitly states 'there was no significant difference in the performance between the two models.' Thus, the claimed multiphase advantage is not supported by any statistically significant evidence. Both models significantly improve over LQ input, meaning the observed gains could come from the network's ability to invert the synthetic degradation rather than from the multiphase fusion. Since the abstract and title emphasize 'leveraging multiphase CT,' this missing comparison is the load-bearing gap. The synthetic-degradation fidelity concern raised by the reader is an external-validity issue, but it does not address whether the core mechanism (multiphase fusion) works at all; the latter is more fundamental and internally testable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D progressive fusion and non-local (PFNL) network that takes three degraded CT phases (non-contrast, arterial, portal venous) as input and produces an enhanced portal venous volume, trained on the VinDr-Multiphase dataset with synthetic multi-stage degradation. The authors evaluate the method with PSNR/SSIM and a proxy task: pancreas segmentation by TotalSegmentator on the enhanced portal venous scans. They report that the proposed 3D-PFNL improves pancreas segmentation Dice from 68.9±19.3% on low-quality input to 71.7±15.6% and NSD from 22.0±8.0 to 25.0±8.5, with paired Wilcoxon p-values of .034 and .025, respectively. The paper claims to be the first to harness multiphase CT for scan quality enhancement and improved pancreas segmentation.","tokens_in":7742,"tokens_out":3396,"duration_ms":31902,"significance":"If the multiphase fusion claim were robustly supported, the work would be a useful proof-of-concept for using complementary CT phases to improve image quality and downstream automated segmentation. The strengths include the use of a public multiphase dataset, an external segmentation tool for the proxy evaluation, paired nonparametric statistical tests, and an explicit acknowledgment of the small sample and synthetic-degradation limitations. However, the central novelty—that multiphase information adds value over single-phase enhancement—is not statistically supported by the reported experiments, and the synthetic degradation setup limits external validity. As a preliminary study the result is plausible, but the current evidence does not justify the abstract's and title's emphasis on the multiphase advantage.","major_comments":[{"comment":"The central claim that leveraging multiphase CT is advantageous is not supported by the reported comparisons. Table I shows 3D-PFNL has lower PSNR/SSIM than the single-phase 3D-RCAN baseline (28.16/0.8733 vs 28.77/0.8794), and Table II shows only a 0.9 percentage-point Dice and 1.0 percentage-point NSD improvement over 3D-RCAN. The Discussion explicitly states that 'there was no significant difference in the performance between the two models.' Since the only significant differences are versus the low-quality input, and 3D-RCAN also achieves significant improvement over that baseline, the observed gains could be attributed to the network's ability to invert the synthetic degradation rather than to multiphase fusion. Please provide a direct statistical comparison between 3D-PFNL and 3D-RCAN, or revise the title/abstract to present this as a preliminary single-arm proof-of-concept without claiming a multiphase advantage.","section":"III.A, Table I; III.B, Table II; IV. Discussion"},{"comment":"The load-bearing premise that the synthetic multi-stage degradation faithfully represents real clinical low-quality CT is neither validated nor tested. The model is trained and evaluated on the same VinDr dataset, with low-quality inputs generated by blur, Gaussian/Poisson noise, resizing at scale factor 4, and second-order Real-ESRGAN-style degradation. The authors acknowledge in the Discussion that generalizability to external datasets such as low-dose CT is undetermined, but this is not merely a future direction: it is central to the clinical utility claim. A validation on real low-dose, motion-degraded, or artifact-affected CT scans, or at least an external dataset with naturally degraded images, is needed to establish that the measured segmentation gains are not an artifact of the synthetic degradation model.","section":"II.C; IV. Discussion"},{"comment":"The statistical reporting is insufficient to support the strength of the conclusions. The test set contains only n=16 scans, and the reported p-values (.034 for Dice, .025 for NSD) are borderline; with one or two cases changed, these would not reach significance. In addition, Table I reports PSNR/SSIM as point estimates without variance, confidence intervals, or any significance test, so the claim that both models 'enhanced the quality' of the low-quality input is not statistically quantified. Please report per-case results, confidence intervals, or effect sizes, and state explicitly which comparisons the p-values in Fig. 3 correspond to.","section":"III.A and III.B; Tables I and II"}],"minor_comments":[{"comment":"The text uses '3D-RFNL' where it should read '3D-PFNL'; please correct this typo.","section":"III.B, first sentence"},{"comment":"The p-values shown in the box plots are not identified in the caption; specify the statistical test and the exact pairwise comparisons for each value.","section":"Fig. 3 caption"},{"comment":"For reproducibility, the degradation procedure should specify the ranges of the Gaussian blur kernel sizes, noise levels, and the order of the second-order degradation operations rather than saying only that a random combination was chosen.","section":"II.C"},{"comment":"The edge-loss weight lambda is said to be 'empirically set to 0.7 based on prior work,' but no reference or ablation is provided; please cite the prior work or report a sensitivity analysis.","section":"II.D"},{"comment":"The reference standard construction relies on an automatic model followed by manual correction; please state how much of the annotation was corrected and whether the radiologist was blinded to the model outputs.","section":"II.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a compact proof-of-concept with a plausible idea, but the primary novelty claim is not statistically established against the single-phase baseline. The small test set and synthetic degradation are acknowledged by the authors, yet the title and abstract still assert a multiphase benefit. I would support a major revision that either adds a direct multiphase-vs-single-phase statistical comparison and external validation, or substantially tones down the claim to a feasibility study. The paper may be better suited to a venue that emphasizes preliminary clinical imaging results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clear, honest proof-of-concept that fails to demonstrate its own headline claim. The 3D-PFNL network, which fuses non-contrast, arterial, and portal venous phases, is not significantly better than 3D-RCAN, a single-phase baseline, on either image quality or the segmentation proxy. The Discussion says exactly that: \"there was no significant difference in the performance between the two models.\" So the claim that leveraging multiphase CT helps is not supported by the data. The observed gains over the low-quality input could just come from the network learning to invert the synthetic degradation.\n\nWhat the paper does well: it extends the 2D PFNL video super-resolution model to 3D, adds a Sobel edge loss, and evaluates the enhancement through a downstream pancreas segmentation task using TotalSegmentator. The authors are transparent about the small test set, the synthetic degradation, and the lack of external validation. They also cite prior multiphase cardiac CT denoising (Sandfort et al.), which undercuts their \"first to harness multiphase CT\" claim.\n\nThe soft spots are not minor. The missing ablation is the load-bearing one: there is no single-phase version of 3D-PFNL, so the comparison to 3D-RCAN mixes architecture and input-phases. Also, 3D-PFNL actually has lower PSNR/SSIM than 3D-RCAN (28.16/0.8733 vs 28.77/0.8794), and the segmentation gains are a few percentage points with overlapping confidence intervals. The synthetic degradation is a separate external-validity issue; the authors acknowledge it, but it means the core mechanism is only tested on artificially degraded data from the same dataset. The test set is n=16, and code and data are not provided.\n\nWho is this for? Readers working on CT quality enhancement or multiphase imaging might find it a useful proof-of-concept, especially the edge-loss and proxy-evaluation design. But the central claim needs a proper single-phase ablation and ideally real low-dose or artifact-corrupted data before it can be believed.\n\nRecommendation: send it to peer review, but with the expectation of major revision. The idea is worth testing, and the paper is honest, but the current evidence does not support the multiphase advantage.","headline":"An honest proof-of-concept that never actually shows the multiphase fusion is what helps; the paper admits the multiphase model is not significantly better than a single-phase baseline, so the headline claim is unsupported as it stands.","tokens_in":8270,"tokens_out":1499,"would_cite":false,"duration_ms":15127,"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":"Multiphase CT fusion restores portal venous quality and lifts pancreas segmentation Dice by about 3 points.","keywords":["multiphase CT","portal venous phase","CT quality enhancement","pancreas segmentation","progressive fusion network","non-local network","synthetic degradation","computed tomography"],"falsifier":"Run the same 3D-PFNL model on real low-dose or artifact-corrupted portal venous CT with paired high-quality references and show that the Dice gain over the low-quality input disappears or fails to reach significance, which would indicate the synthetic degradation does not capture the real-world failure modes.","tokens_in":7300,"feed_emoji":"🩻","tokens_out":7241,"duration_ms":55542,"temperature":0.7,"pith_summary":"This paper tries to establish that the multiple CT phases already acquired in a routine multiphase study can be reused to improve the quality of one phase instead of being discarded. The authors train a 3D progressive fusion and non-local network on degraded non-contrast, arterial, and portal venous scans to reconstruct a higher-quality portal venous volume, then measure clinical utility through a proxy pancreas segmentation task. They report that the restored portal venous phase improves pancreas segmentation by about three Dice points (68.9% to 71.7%) and by a similar margin in normalized surface distance, both statistically significant. If the claim holds, then complementary contrast and boundary information in existing multiphase CT is a free resource for scan-quality restoration and downstream organ segmentation.","feed_headline":"Three CT phases fuse to lift pancreas segmentation by ~3%","feed_subtitle":"Restored portal venous scans raise Dice from 68.9% to 71.7% over low-quality input.","key_machinery":"The central object is the 3D progressive fusion and non-local (PFNL) network, a video super-resolution architecture extended to three dimensions and repurposed to take three co-registered low-quality CT phases as input and output one high-quality portal venous volume. Its non-local component lets the network borrow contrast and boundary information across phases, while progressive fusion integrates the phases at multiple scales. Training uses an L1 reconstruction loss together with a 3D Sobel edge-based loss that biases the network toward preserving organ boundaries, and a second-order synthetic degradation pipeline generates the low-quality inputs from the original portal venous scans. The restored portal venous volume is then fed to a public multi-organ segmentation tool as a proxy readout of clinical utility.","core_discovery":"On the paper's own terms, the central discovery is that three co-registered CT phases—non-contrast, arterial, and portal venous—can be fused by a 3D progressive fusion and non-local network to restore a degraded portal venous scan, and that this restoration carries through to a downstream task. On the test set, the restored portal venous phase raises the Dice similarity coefficient from 68.9 ± 19.3% for low-quality input to 71.7 ± 15.6% (p = .034) and normalized surface distance from 22.0 ± 8.0% to 25.0 ± 8.5% (p = .025). The improvement over low-quality input is statistically significant; the difference against a single-phase super-resolution baseline is not, though the multiphase model yields sharper pancreatic boundaries and a lower Dice standard deviation. The authors present this as the first demonstration that multiphase CT can be harnessed for scan quality enhancement.","pith_inferences":["A natural next step is to test the pipeline on real low-dose or motion-corrupted CT with paired high-quality references, since the current evidence comes entirely from synthetic degradation and transfer to real artifact patterns is unproven.","The non-local multiphase fusion idea may transfer to other imaging settings where complementary volumes are routinely acquired, such as MRI contrasts or dual-energy CT.","An ablation study that removes the 3D Sobel edge loss would isolate how much of the segmentation gain is due to the edge-preserving term versus the multiphase fusion itself.","The artifact the authors note—enhanced aorta contrast appearing in the restored portal venous phase—would need monitoring in any clinical use to ensure the network does not introduce features from other phases that mislead interpretation."],"forward_implications":["Restoring the portal venous phase with multiphase information recovers part of the segmentation loss caused by CT degradation, raising Dice from 68.9% to 71.7% and normalized surface distance from 22.0% to 25.0%.","The multiphase approach narrows the gap in pancreas segmentation between degraded and original scans, as shown by the lower Dice standard deviation (15.6 versus 19.3) and by one test case where segmentation failed on low-quality input but succeeded after restoration.","Because the method is not tailored to the pancreas, the same enhanced portal venous volume could improve segmentation of other abdominal structures without retraining for each organ.","The framework can in principle be applied to enhance other phases, such as non-contrast CT, which is used for opportunistic screening.","Quality metrics such as PSNR and SSIM do not by themselves predict downstream segmentation benefit, so a proxy task is needed to judge the restoration's clinical value."],"supporting_citations":[{"why":"Supplies the progressive fusion and non-local architecture that the paper extends from video super-resolution to 3D multiphase CT.","marker":"[19]"},{"why":"Provides the public segmentation tool used in the proxy task to measure the downstream utility of the restored portal venous phase.","marker":"[20]"},{"why":"Source of the 168 multiphase abdominal CT studies used for training, validation, and testing.","marker":"[21]"},{"why":"Provides the initial automatic pancreas annotations that a radiologist manually corrected to build the reference standard.","marker":"[22]"},{"why":"Registration method used to align the non-contrast and arterial phases to the portal venous reference volume.","marker":"[23]"},{"why":"Defines the second-order degradation process used to synthesize low-quality CT inputs from high-quality originals.","marker":"[24]"},{"why":"Single-image super-resolution baseline compared against the 3D-PFNL model on restoration quality and downstream segmentation.","marker":"[25]"}],"fun_headline_variants":["Multiphase CT fusion boosts pancreas segmentation by 3%","3 CT phases restore portal venous scans for better pancreas Dice","Fusing three CT phases lifts pancreas segmentation Dice","Multiphase CT quality boost aids pancreas segmentation","Portal venous CT restored via three-phase fusion, aids pancreas"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic multi-stage degradation used to produce low-quality training and test scans faithfully represents real clinical low-quality CT, because the model is trained and evaluated on the same dataset under that simulation.","fun_headline_variants_meta":{"raw":{"variants":["Multiphase CT fusion boosts pancreas segmentation by 3%","3 CT phases restore portal venous scans for better pancreas Dice","Fusing three CT phases lifts pancreas segmentation Dice","Multiphase CT quality boost aids pancreas segmentation","Portal venous CT restored via three-phase fusion, aids pancreas"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1435,"prompt_tokens":944,"completion_tokens":491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":413}},"tokens_in":560,"tokens_out":491,"duration_ms":4119,"temperature":1.0,"reasoning_tokens":413,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:28:17.467701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 3D-PFNL model on real low-dose or artifact-corrupted portal venous CT with paired high-quality references and show that the Dice gain over the low-quality input disappears or fails to reach significance, which would indicate the synthetic degradation does not capture the real-world failure modes.","supporting_citations":[{"cited_title":"Progressive Fusion Video Super-Resolution Network via Exploiting Non-Local Spatio-Temporal Correlations,","cited_arxiv_id":null,"evidence_quote":"Supplies the progressive fusion and non-local architecture that the paper extends from video super-resolution to 3D multiphase CT."},{"cited_title":"TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images,","cited_arxiv_id":null,"evidence_quote":"Provides the public segmentation tool used in the proxy task to measure the downstream utility of the restored portal venous phase."},{"cited_title":"Phase Recognition in Contrast-enhanced CT Scans Based on Deep Learning and Random Sampling,","cited_arxiv_id":null,"evidence_quote":"Source of the 168 multiphase abdominal CT studies used for training, validation, and testing."},{"cited_title":"A Reproducible Evaluation of ANTs Similarity Metric Performance in Brain Image Registration,","cited_arxiv_id":null,"evidence_quote":"Registration method used to align the non-contrast and arterial phases to the portal venous reference volume."},{"cited_title":"Real-ESRGAN: Training Real- World Blind Super-Resolution with Pure Synthetic Data,","cited_arxiv_id":null,"evidence_quote":"Defines the second-order degradation process used to synthesize low-quality CT inputs from high-quality originals."},{"cited_title":"Image Super- Resolution Using Very Deep Residual Channel Attention Networks,","cited_arxiv_id":null,"evidence_quote":"Single-image super-resolution baseline compared against the 3D-PFNL model on restoration quality and downstream segmentation."}],"review_version":1}