{"id":"17783eaf-8c14-409c-9148-1616294c0ac4","arxiv_id":"2501.01456","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised framework jointly trains low-dose full-view, sparse-view, and limited-view CT reconstruction networks by enforcing cross-task consistency, with reported gains on simulated and real clinical data.","lead":"This paper trains three CT reconstruction networks on low-dose, sparse-view, and limited-view versions of the same sinogram, letting them learn from each other without paired clean images. If the reported gains hold, it offers a way to reduce CT radiation dose without needing matched normal-dose training data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only asymmetric supervision is Eq. (7)'s anchor to the corrupted low-mAs FBP; since Eqs. (5)-(6) are symmetric agreement terms, the objective admits a trivial all-outputs-equal-mu_ld solution, so cross-task mutual learning may not be the active mechanism behind the reported gains.","rationale":"The strongest claim is not just numerical superiority but a mechanism: three tasks improve by learning from each other without clean data. For that mechanism to be true, the mutual losses must provide a gradient signal pointing toward true anatomy. Inspection of Eqs. (5)-(7) shows they cannot: pairwise agreement terms only enforce equality, and the only tie to an observed image is through mu_ld. A trivial fixed point is all outputs equal to mu_ld, which achieves zero total loss whenever f_ld is identity and the other branches are expressive enough to emulate mu_ld from their inputs. In practice the branches cannot perfectly emulate mu_ld, so the system behaves as a denoising auto-encoder with a noisy anchor plus an architectural bottleneck. The reported improvements are consistent with that, but the paper's ablations do not isolate the mutual-learning contribution. A control with direct regression to mu_ld would settle the question: if the control matches SS-CTML, the central claim is overstated; if not, mutual learning adds real value. The reader's weakest assumption about Eq. (7) is related but less structural; my concern is that even a clean anchor would reduce the method to supervised regression, so the mutual losses need independent validation. I would keep the conditional verdict but add this specific control requirement as a condition for acceptance.","tokens_in":16865,"tokens_out":7856,"duration_ms":84801,"concrete_test":"Run a control ablation with identical architectures, data splits, and hyperparameters, but replace L_total in Eq. (8) by L_ctrl = ||mu_prior_ld - mu_ld||^2 + ||mu_out_ld - mu_ld||^2 + ||mu_prior_sv - mu_ld||^2 + ||mu_out_sv - mu_ld||^2 + ||mu_prior_lv - mu_ld||^2 + ||mu_out_lv - mu_ld||^2, i.e., remove Eqs. (5) and (6) and regress every branch directly to the low-mAs FBP. Compare SS-CTML vs. L_ctrl on the 560-slice test set per task with paired Wilcoxon tests. If no significant difference (p >= 0.05) and mean PSNR differences are below 0.1 dB, the cross-task mutual losses are not load-bearing; if SS-CTML is significantly better, the consensus constraint contributes beyond the mu_ld anchor. Also report the distance of each output to mu_ld versus to the normal-dose reference to check whether the method is reproducing the corrupted anchor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that with only low-mAs sinograms, cross-task mutual learning lets the three subnetworks improve all three tasks simultaneously. The load-bearing weakness is in the loss design. Eq. (5) and Eq. (6) are symmetric pairwise L2 agreement terms; their gradients vanish at any consensus configuration. The only asymmetric term is Eq. (7), which anchors the FVCT branch to the low-mAs FBP image mu_ld. Therefore the consensus direction is set entirely by mu_ld. If f_ld learns identity and f_sv/f_lv learn to map their extracted sparse/limited sinograms to mu_ld, then L_total is driven to about zero without recovering normal-dose anatomy. The reported FVCT gain over FBP (39.58 vs 34.34 dB in Table I) shows the trained model is not exactly the identity solution, but that gain then comes from the architecture's implicit regularization rather than from the mutual-learning principle. The paper's ablations (Table IV) remove one task or one module, never the mutual losses while keeping the identical anchor and architecture, so they cannot establish that cross-task consensus, rather than pseudo-label regression to mu_ld with a U-net backbone, is what improves SVCT and LVCT. This is distinct from the stated outcome and is testable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SS-CTML, a self-supervised framework for low-mAs full-view CT (FVCT), sparse-view CT (SVCT), and limited-view CT (LVCT) reconstruction. From a low-mAs full-view sinogram it extracts sparse-view and limited-view sinograms, trains three dual-domain subnetworks with prior neural modules and compensation modules, and optimizes them with pairwise mutual-learning losses on prior images and final outputs plus a reconstruction-consistency loss anchored to the low-mAs FBP image. Experiments on the Mayo dataset at quarter/sixth/eighth dose and on ELCAP real data report PSNR/NMSE/SSIM; SS-CTML outperforms the unsupervised baselines and, for SVCT and LVCT, the supervised baselines, while for FVCT it is below RED-CNN. The authors conclude that cross-task mutual learning improves all three tasks without normal-dose paired training.","tokens_in":17143,"tokens_out":6348,"duration_ms":64187,"significance":"If the central claim holds, SS-CTML would be a practically valuable self-supervised alternative to paired-data reconstruction, since it learns from low-mAs sinograms alone and improves three tasks simultaneously; the paper provides extensive quantitative tables across three dose levels, a real-data demonstration, and an ablation of the main modules. The evaluation is not circular because the normal-dose references are held out from training. However, the proposed mutual-learning mechanism is not isolated by the current experiments, and the baseline comparisons need more transparency; with those gaps closed, the work would be a solid contribution.","major_comments":[{"comment":"The central claim that cross-task mutual learning is the active mechanism is not established by the loss design or the ablations. Lprior_ml and Lout_ml are symmetric pairwise agreement terms whose gradients vanish at any consensus configuration, and Lrc in Eq. (7) is the only asymmetric term, anchoring all outputs to the corrupted FBP image mu_ld. The objective therefore admits a near-zero-loss solution in which f_ld is identity and f_sv/f_lv map their inputs to mu_ld, so the improvements over FBP reported in Tables I-III could in principle come from the U-net backbone's implicit regularization rather than from cross-task consensus. Table IV removes whole tasks or modules (w/o FVCT, w/o SVCT, w/o LVCT, w/o PNM, w/o DDNM) but never removes the mutual-learning losses while keeping the same anchor and architecture; notably, w/o FVCT drops SVCT PSNR to 29.10 dB, which is consistent with losing the Lrc anchor rather than with losing mutual learning. Please add an ablation training the same three subnetworks with only Lrc (or only per-task consistency toward mu_ld) and report the three tasks, and specify which loss terms remain in the w/o FVCT variant.","section":"III-C, Eqs. (5)-(8); V-C, Table IV"},{"comment":"The quantitative comparison to supervised baselines is currently difficult to interpret because the training protocols of the competing methods are not specified. It is not stated whether RED-CNN, FBPConvNet, iRadonMap, FreeSeed, Sam's Net, Dn-Dp, and DOLCE were retrained on the same simulated low-dose data with the same view geometries, or with which hyperparameters and implementations. The very low DOLCE scores (18.84/18.29/19.03 dB across dose levels versus 25-34 dB for other LVCT methods) suggest the baseline may not have been adapted to the mixed limited-view and low-mAs setting, which would bias the comparison in favor of SS-CTML. Please provide the training protocol for each baseline and report paired statistical significance tests (for example, Wilcoxon signed-rank or paired t-test over slices) for the main comparisons.","section":"IV-C, Tables I-III"},{"comment":"The framework's assumption that the three FBP images contain complementary but consistent anatomy so that consensus via Eqs. (5)-(6) drives toward true anatomy is not directly tested. Since Eq. (7) anchors the entire objective to the low-mAs FBP image mu_ld, a heavily corrupted mu_ld could make the consensus a degraded average; the paper does not report how close the converged consensus is to the normal-dose reference on a case-by-case basis, or compare with an alternative anchor such as a filtered or iterative reconstruction of pld. An additional experiment on a numerical phantom, or with a deliberately corrupted mu_ld, would show whether mutual learning recovers truth rather than a common artifact pattern, and would directly address the load-bearing assumption.","section":"III-A, Eq. (7); V-A"}],"minor_comments":[{"comment":"Section I contains the typo 'real clinal practices' and should read 'real clinical practices'; Section VII contains 'the the effectiveness' and should read 'the effectiveness'.","section":"I, VII"},{"comment":"The caption says 'SVCT reconstruction task' but the figure shows LVCT results; the caption should read 'LVCT reconstruction task'.","section":"Fig. 9"},{"comment":"The phrase 'the batch number was 1000' is ambiguous; please specify the batch size, the number of training epochs, and the learning-rate schedule, and justify the unusual Adam beta values (0.5, 0.9).","section":"IV-B"},{"comment":"The axis label 'Epoches' should be 'Epochs', and the legends for the three tasks in panel (b) are difficult to read; please enlarge the fonts and clarify the panel labels.","section":"Fig. 3"},{"comment":"The use of concatenation (the symbol ⊕) only for the FVCT branch, while SVCT and LVCT use masking, is not explained; please clarify how concatenation acts as sinogram compensation in the low-mAs case.","section":"Eq. (3)"},{"comment":"The paper does not release code or data; adding a reproducibility statement and the trained models would strengthen the empirical claims.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper is within scope and the empirical results are promising, but the main mechanism claim is under-supported by the current ablations; the trivial consensus solution to Eqs. (5)-(8) is a serious correctness risk that needs to be addressed experimentally before acceptance. I would also encourage the editor to check whether the extremely low DOLCE numbers reflect a mismatch in experimental setup, since otherwise the LVCT comparison may be misleading."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of SS-CTML. The core idea is genuinely nice: from a single low-dose full-view sinogram, you derive sparse-view and limited-view tasks, then train three networks jointly with mutual consistency losses, so no paired normal-dose data is needed. The reported numbers are strong—SVCT PSNR 39.25 dB at quarter-dose, beating supervised FBPConvNet and FreeSeed, and LVCT 33.92 dB, close to Sam's Net—and the architecture (prior module, forward projection compensation, dual-domain refinement) is sensible. The real-data demonstration, though qualitative, is a plus.\n\nBut the central mechanistic claim is underdetermined by the loss design. Eq. (7) anchors only the FVCT branch to the low-mAs FBP image μ_ld. Eqs. (5)-(6) are symmetric pairwise agreement terms; they vanish at any consensus. So the objective admits a trivial solution where all three outputs equal μ_ld, and more generally, the anchor—not any genuine mutual exchange—sets the direction. The ablations remove one task or one module, but never the mutual losses while keeping the anchor and the same backbone. A reader cannot tell whether the SVCT/LVCT gains come from cross-task consensus or simply from pseudo-label regression to μ_ld with a U-net that has good implicit regularization. The w/o FVCT row (SVCT drops to 29.1 dB) shows the anchor is load-bearing, which is consistent with the stress-test reading.\n\nThere are smaller issues: no code or data release, no significance testing, no training details for the baselines, and DOLCE on LVCT looks set up to fail because the test data mixes limited-view and low-mAs artifacts—that comparison should be rerun or contextualized. The discussion's suggestion that this could become a foundation model outruns the evidence.\n\nNone of this means the paper is wrong. The improvement over the trivial solution is real, which suggests the architecture is doing something useful. But the stated contribution—that mutual learning across tasks is the active mechanism—needs a missing control experiment: train all three branches with L_rc anchored to μ_ld (or to each branch's own FBP) and no mutual losses; or add mutual losses on top of that and compare. That single experiment would settle it.\n\nI'd send this to review: the idea is practical, the experiments are extensive, and the flaw is addressable with an ablation rather than fatal. If the control confirms the mutual loss helps, this is a worthy contribution to the self-supervised CT literature. If not, the paper still has a useful multi-task setup, but the claims need to be toned down.","headline":"Clever self-supervised multi-task CT setup with strong numbers, but the loss design leaves the mutual-learning mechanism underdetermined; worth a serious look with a requested control ablation.","tokens_in":17692,"tokens_out":5155,"would_cite":false,"duration_ms":48045,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Three CT reconstruction tasks train each other without any normal-dose paired data.","keywords":["CT image reconstruction","self-supervised learning","cross-task mutual learning","sparse-view CT","limited-view CT","low-dose CT","dual-domain network","knowledge distillation"],"falsifier":"Feed the framework a test slice in which the low-mAs FBP image carries a strong known bias that the sparse- and limited-view degradations do not share; if the mutual-learning outputs reproduce that biased structure instead of removing it, the reconstruction-consistency anchor is pulling the consensus toward the artifact rather than toward true anatomy.","tokens_in":16664,"feed_emoji":"🩻","tokens_out":6755,"duration_ms":61856,"temperature":0.7,"pith_summary":"This paper tries to establish that low-dose full-view CT, sparse-view CT, and limited-view CT reconstruction can be improved simultaneously by a self-supervised mutual-learning framework that never uses normal-dose images as training targets. The authors derive sparse-view and limited-view sinograms from a single low-dose full-view scan, build three subnetworks for the three tasks, and constrain the networks to agree on shared anatomy. The payoff, if the claim holds, is that routine low-dose scans alone can train networks for several CT artifact-removal tasks, removing a major obstacle to clinical use of supervised deep learning, which is hard to obtain in practice.","feed_headline":"Three CT tasks train each other without normal-dose data","feed_subtitle":"Mutual-learning losses let one low-dose scan teach three reconstruction networks, beating several supervised baselines.","key_machinery":"The load-bearing machinery is a trio of subnetworks, one per task, each made of a prior neural module (PNM) and a dual-domain neural module (DDNM). The PNM takes the contaminated FBP image and produces a cleaned prior image; that prior is forward-projected into the sinogram domain, where for sparse- and limited-view tasks a binary mask fills the missing angular positions with the forward projection of the prior while retaining measured data elsewhere, and for the full-view task the projection is concatenated with the measured sinogram. The compensated sinogram then passes through the DDNM, which is a sinogram network, a back-propagatable FBP layer, and an image network. The cross-task mutual-learning losses $L_{\\text{prior}}^{\\text{ml}}$ and $L_{\\text{out}}^{\\text{ml}}$ are $\\ell^2$ distances among the three priors and among the three final outputs, and the reconstruction-consistency loss $L_{\\text{rc}}$ ties the full-view branch to the low-mAs FBP image $\\mu_{ld}$; the total loss is their sum. These objectives are what let one low-dose dataset supervise three tasks at once.","core_discovery":"The central claim is that the FBP reconstructions of the same slice from low-mAs full-view, sparse-view, and limited-view scans carry complementary information about true anatomy, so forcing three subnetworks to produce mutually consistent priors and final outputs, while anchoring the full-view branch to the low-mAs FBP image, optimizes all three tasks without any normal-dose reference. On simulated quarter-, sixth-, and eighth-dose data from a public dataset, the paper reports that SS-CTML outperforms the supervised baselines FBPConvNet, iRadonMap, and FreeSeed for sparse-view CT, outperforms FBPConvNet, Sam's Net, and DOLCE for limited-view CT, and is competitive with the supervised RED-CNN and diffusion-based Dn-Dp for low-mAs full-view CT. On real clinical low-dose scans, the trained model reduces noise and streak artifacts qualitatively for all three tasks, with no quantitative reference available.","pith_inferences":["I would expect the same mutual-agreement recipe to extend beyond CT geometries: any set of image-degradation tasks derived from one measurement, such as different sparse-view rates or masked sinogram regions, could be jointly self-supervised, with retraining per geometry as the paper acknowledges.","The weakest spot I see is the anchor: if the full-view low-dose FBP image itself contains a structured artifact, Eq. (7) gives that artifact a privileged position in the consensus, so I would test the method under extreme noise or with a deliberately biased reference.","One network with task-specific heads might replace the three separate subnetworks and preserve most of the mutual-learning benefit while cutting memory; the paper itself flags the three-network design as redundant.","Because the three tasks share only the underlying anatomy, the mutual losses act as a form of self-distillation across geometries; comparing against a single-network denoiser trained with the same total compute would isolate what the cross-task coupling actually buys."],"forward_implications":["A single self-supervised training run on low-dose sinograms produces networks for all three reconstruction tasks at once, reducing the need for paired normal-dose training data in each task.","Per the paper's results, the sparse-view and limited-view tasks can match or exceed several dedicated supervised methods, so self-supervised mutual learning is a viable alternative where paired data are unavailable.","The ablation study implies that each task contributes: removing the full-view task sharply degrades the remaining two, so the cross-task correlation is the source of the gain, not a side effect.","The trained model transfers to real clinical low-dose scans of different anatomy and reduces visible noise and artifacts, suggesting the learned consensus is not overfit to the simulated training geometry.","Both the prior-image consensus and the final-output consensus contribute, since ablating either the PNM or the DDNM lowers performance; the dual-domain structure carries part of the benefit."],"supporting_citations":[{"why":"Supplies the mutual-learning idea that peer networks train each other without a teacher network.","marker":"[29]"},{"why":"Shows self-supervised denoising from unpaired sinogram data, which the framework builds on.","marker":"[11]"},{"why":"Provides the dual-domain Radon-inversion design and serves as a supervised baseline for FVCT and SVCT.","marker":"[33]"},{"why":"The self-augmented limited-angle network used as the LVCT supervised baseline the method compares against.","marker":"[4]"},{"why":"The supervised low-dose CT denoiser used as the FVCT baseline.","marker":"[35]"},{"why":"The FBPConvNet baseline for SVCT and LVCT image postprocessing.","marker":"[3]"},{"why":"The sparse-view CT baseline that the method outperforms across dose levels.","marker":"[36]"},{"why":"The zero-shot diffusion baseline for low-dose FVCT denoising.","marker":"[22]"},{"why":"The limited-angle diffusion baseline for LVCT reconstruction.","marker":"[17]"}],"fun_headline_variants":["Self-supervised CT trio learns from one low-dose scan","Mutual learning lifts CT reconstruction without normal-dose pairs","Cross-task mutual learning boosts CT without high-dose data","Low-dose CT teaches three networks via mutual consistency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the three degraded FBP images of the same slice are complementary but consistent views of true anatomy, so that forcing them to agree pushes each network toward a clean image; in particular, every task is anchored to the low-mAs FBP image $\\mu_{ld}$, and if that image is too corrupted the mutual-learning consensus could converge to a degraded average rather than the truth.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised CT trio learns from one low-dose scan","Mutual learning lifts CT reconstruction without normal-dose pairs","Cross-task mutual learning boosts CT without high-dose data","Low-dose CT teaches three networks via mutual consistency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1259,"prompt_tokens":982,"completion_tokens":277,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":214}},"tokens_in":598,"tokens_out":277,"duration_ms":3328,"temperature":1.0,"reasoning_tokens":214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:54:33.650742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the framework a test slice in which the low-mAs FBP image carries a strong known bias that the sparse- and limited-view degradations do not share; if the mutual-learning outputs reproduce that biased structure instead of removing it, the reconstruction-consistency anchor is pulling the consensus toward the artifact rather than toward true anatomy.","supporting_citations":[{"cited_title":"Deep mutual learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-learning idea that peer networks train each other without a teacher network."},{"cited_title":"Radon Inversion via Deep Learning,","cited_arxiv_id":null,"evidence_quote":"Provides the dual-domain Radon-inversion design and serves as a supervised baseline for FVCT and SVCT."},{"cited_title":"Sam’s net: a self-augmented multistage deep-learning network for end-to-end recon- struction of limited angle CT,","cited_arxiv_id":null,"evidence_quote":"The self-augmented limited-angle network used as the LVCT supervised baseline the method compares against."},{"cited_title":"Low-dose CT with a residual encoder-decoder convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"The supervised low-dose CT denoiser used as the FVCT baseline."},{"cited_title":"Deep con- volutional neural network for inverse problems in imaging,","cited_arxiv_id":null,"evidence_quote":"The FBPConvNet baseline for SVCT and LVCT image postprocessing."},{"cited_title":"FreeSeed: Frequency- band-aware and self-guided network for sparse-view CT reconstruc- tion,","cited_arxiv_id":null,"evidence_quote":"The sparse-view CT baseline that the method outperforms across dose levels."},{"cited_title":"DOLCE: A model-based probabilistic diffusion framework for limited-angle CT reconstruction,","cited_arxiv_id":null,"evidence_quote":"The limited-angle diffusion baseline for LVCT reconstruction."}],"review_version":1}