{"id":"e62ed20f-7bd3-4ef5-9a23-be20e86cf031","arxiv_id":"2506.04116","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"TSSC-Net uses a diffusion model conditioned on start and end frames to generate 6x more MRI time frames, then a tri-directional Mamba network to fix spatial inconsistencies.","lead":"This paper proposes TSSC-Net, a two-stage deep learning framework that generates missing time frames in 4D MRI scans and then corrects cross-slice inconsistencies. It reports better interpolation quality than registration-based baselines, especially on large-motion knee imaging, which could reduce scan time while preserving dynamic detail.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage 1's diffusion target is written as x0 = I1, the same frame used as a conditioning input in Eq. (2); taken literally, the model reconstructs a conditioning frame rather than generating intermediate frames, which would invalidate the 6x interpolation claim.","rationale":"I read the paper in good faith and identified what the central claim requires: Stage 1 must synthesize frames that are not among its conditioning inputs, and Stage 2 must correct cross-slice inconsistencies without simply memorizing artifact patterns. The reader's weakest assumption concerned Stage 2's unspecified pairing procedure and the risk of memorization. That is a legitimate concern, but I found a more immediate problem in Stage 1: Section 2.2 explicitly writes x0 = I1 while Eq. (2) conditions on I0 and I1, so the target and a condition coincide. If this is literal, the model is trained to reconstruct one of the input frames, and the '6x temporal super-resolution' claim is unsupported regardless of Stage 2. A charitable correction would replace the target with a middle frame, but the paper does not provide the corrected formulation, the training objective, or the sampling procedure. The 'single inference step' claim is also undefined because DDIM acceleration is mentioned without a step count, and standard DDIM is an iterative sampler. These are addressable issues, but until the authors specify the diffusion target and demonstrate that the outputs are not copies of a conditioning frame, the central claim cannot be assessed. I therefore recommend moving from CONDITIONAL to UNVERDICTED: the current text contains an internal contradiction at the core of the method, and the missing details prevent a fair conditional acceptance.","tokens_in":6892,"tokens_out":4793,"duration_ms":47484,"concrete_test":"Inspect the released code's Stage 1 data loader and training loop: record, for one training sample, the conditioning frames and the noised target x0. If x0 is the same tensor as the end-frame condition I1, the training objective is degenerate and the central claim fails. If x0 is an intermediate frame, run inference with only I0 and I1 as inputs and verify the output is not a near-copy of I1; also report the exact DDIM step count used to support the 'single inference step' claim.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The most load-bearing weakness is in Section 2.2. The paper first defines the start and end frames as I0 and I1, then states 'the forward process is applied to the target frame I1' with x0 = I1. Yet the reverse process in Eq. (2) conditions on I0 and I1, i.e., on the very frame being generated. If this is literal, Stage 1 is not temporal super-resolution at all: it is a conditional reconstruction of an input frame. No intermediate frame is ever produced, so the central claim that TSSC-Net 'generates a complete 12-frame 4D dynamic MRI sequence from only two input frames' collapses. A charitable reading is that x0 should denote an intermediate frame and I1 in Eq. (2) should denote the end-frame condition, but the manuscript never states which frame is noised, which frame is predicted, or how the 10 intermediate frames are obtained in 'a single inference step.' This is not a cosmetic typo: it determines whether the method performs frame interpolation or conditional autoencoding. The reader's concern about Stage 2 training pairs is real, but Stage 1's target/conditioning collision is more fundamental, since a degenerate Stage 1 cannot be rescued by any post-processor.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents TSSC-Net, a two-stage framework for 4D MRI temporal super-resolution. Stage 1 uses a diffusion model conditioned on start and end frames to generate intermediate 2Dt frames, and Stage 2 uses residual tri-directional Mamba blocks to improve 3D spatial consistency. The method is evaluated on the ACDC cardiac MRI dataset and an in-house dynamic knee dataset, with PSNR/SSIM/MAE comparisons against three baselines (VM, TM, UVI-Net) and an ablation that removes Stage 2. The central claims are 6x temporal super-resolution (12 frames from 2 input frames) in a single inference step, and improved performance on large-motion knee imaging (PSNR 20.041 dB vs. 17.580 dB for TM).","tokens_in":7126,"tokens_out":4755,"duration_ms":42915,"significance":"The application is timely, and the reported results on the knee dataset are promising: the PSNR gain of about 2.5 dB over TM and SSIM of 0.730 vs. 0.635 suggest that the method may handle large inter-frame motion better than registration-based interpolation. The manuscript also includes an ablation study, qualitative comparisons, and a commitment to release code, which are strengths. However, the diffusion formulation in Stage 1 is internally inconsistent as written, so the central interpolation claim is not yet established. The paper needs clarification of the generation target, conditioning variables, and inference protocol before its contributions can be assessed.","major_comments":[{"comment":"The forward process is described as being applied to the target frame I1 with x0 = I1, but the reverse process in Eq. (2) conditions on I0 and I1, i.e., on the very frame being generated. Taken literally, Stage 1 reconstructs a conditioning frame rather than synthesizing intermediate frames, which would invalidate the claimed 6x temporal super-resolution. The authors must state explicitly which frame is noised, which frames are used as conditioning, and how the 10 intermediate frames are produced (e.g., one denoising trajectory per intermediate frame, or a single trajectory that outputs all frames). As written, the method is indistinguishable from conditional autoencoding of an input frame.","section":"Section 2.2, Eqs. (1)-(2)"},{"comment":"The repeated claim of \"6x temporal super-resolution in a single inference step\" is inconsistent with the described iterative DDIM sampling: the reverse diffusion process requires multiple denoising steps. The authors need to define the inference procedure: how many DDIM steps are used, whether each output frame is sampled independently, and whether \"single inference step\" means one model forward pass per frame or one complete sampling trajectory. Without this, the computational claim and the temporal-interpolation mechanism remain ambiguous.","section":"Section 2.2 and Section 3.2"},{"comment":"The training data for the spatial consistency enhancement network is not specified. The loss in Eq. (5) is defined on the output of the network, but the paper does not state how input-output pairs are constructed: are the inputs the Stage-1 generated volumes and the targets the ground-truth volumes? If Stage 2 is trained only on the artifacts of Stage 1 on the training set, the ablation in Table 2 may reflect memorization of those specific artifact patterns rather than a generalizable correction capability. The authors should describe the pairing procedure and any data augmentation, and ideally evaluate Stage 2 on held-out volumes.","section":"Section 2.3 and Section 3.2"},{"comment":"The evaluation protocol is under-specified: it is not stated whether the reported PSNR/SSIM are computed on all 12 frames of the generated sequence, only the 10 interpolated frames, or the 2 input frames plus 10 generated frames. This matters because the compared baselines (VM, TM, UVI-Net) are registration-based methods whose outputs may be evaluated on the same frame set. Please specify the exact frames used for evaluation and the preprocessing applied to the in-house knee dataset.","section":"Section 3.3, Table 1"}],"minor_comments":[{"comment":"The text uses \"epsilon_t\" for the fixed variance, but Eq. (2) writes sigma_t^2 I; please make the notation consistent.","section":"Section 2.2, Eq. (2)"},{"comment":"There is a typo: \"de-sign\" should be \"design\".","section":"Section 2.2"},{"comment":"The paper lists \"PyTorch 11.7\"; this is not a standard PyTorch version. It should be \"PyTorch 1.11\" or \"PyTorch 2.x\".","section":"Section 3.2"},{"comment":"The terms \"2Dt\" and \"3D\" are used without definition; please define them at first use to clarify that Stage 1 operates on 2D slices across time and Stage 2 on 3D volumes.","section":"Section 2.1 and Section 2.2"},{"comment":"The labels \"Pred 3Dt\" and \"Pred 3D0\" in Fig. 1 are unclear; please explain what t and 0 denote.","section":"Fig. 1 caption"},{"comment":"The sentence \"To solve temporal fidelity under fast motion\" should be \"To improve temporal fidelity under fast motion\".","section":"Abstract"},{"comment":"The code URL is split across lines in the abstract; please format it as a single hyperlink.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The primary reason for major revision is the Stage 1 formulation, not the empirical claims. If the authors can clarify the notation and provide the missing implementation details for Stage 2, the paper could become acceptable. I would also flag that the in-house knee dataset is not publicly available, so the reported improvement should be independently reproducible at least through the released code; the editor may wish to verify code and data availability before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing. The framework is a two-stage pipeline: a diffusion model interpolates 2D slices from start/end frames, then a residual tri-directional Mamba network cleans up cross-slice inconsistencies in the reassembled 3D volume. The second stage is the more interesting piece, and the knee results are real: 20.04 dB PSNR versus 17.58 for TransMorph, with the ablation dropping to 18.13 without the spatial consistency module. On ACDC, TSSC-Net essentially ties UVI-Net, so the large-motion knee dataset is the evidence that matters.\n\nWhat is genuinely new: the combination of diffusion-based temporal interpolation with a Mamba module that scans along three orthogonal axes for volumetric coherence. I have not seen that specific pairing in the cited literature, and the ablation supports that the tri-directional scan contributes something.\n\nThe soft spots. First, the diffusion formulation in Section 2.2 is wrong as written. Eq. (1) sets x0 = I1, and Eq. (2) conditions on I0 and I1 for every reverse step. Literally, that means the model is reconstructing one of its own conditioning inputs, not producing an intermediate frame. The intended reading is presumably that x0 is an intermediate frame and I0 and I1 are the boundary conditions, but a referee cannot tell from the text. This is the load-bearing issue and it must be fixed.\n\nSecond, the 'single inference step' wording conflicts with DDIM acceleration, which normally uses multiple sampling steps. The authors should state exactly how many steps they run.\n\nThird, no diffusion-based baseline is included, even though LDMVFI and VDT are cited. That leaves the central comparison to registration-based methods and one unsupervised flow approach.\n\nFourth, the Stage 2 training procedure is underspecified. How are the input volumes to the spatial consistency network constructed during training? If they are Stage 1 outputs, fine, but that is not said; if they are ground-truth volumes, then the post-processor is being trained on a different distribution from what it sees at inference.\n\nFifth, the knee dataset is 12 in-house cases, private, with no significance tests, and the code is not released despite the abstract promising it. That is a lot to rest on.\n\nFor peer review: I would send it out, with a firm request to rewrite the diffusion equations, report sampling steps, add a diffusion baseline, specify Stage 2 pairing, and make at least some data or code available. The core idea is worth testing, but the current manuscript is not self-contained enough to verify the central claim.","headline":"A plausible two-stage diffusion+Mamba pipeline for 4D MRI interpolation with real gains on large-motion knee data, but the diffusion target notation needs correction before the 6x interpolation claim is trustworthy.","tokens_in":7691,"tokens_out":3204,"would_cite":false,"duration_ms":29236,"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":"TSSC-Net claims that conditioning a diffusion model on the start and end frames of a 4D MRI scan generates all intermediate frames (6x temporal super-resolution in a single step), and a tri-directional Mamba post-processor restores…","keywords":["4D MRI","temporal super-resolution","diffusion model","spatial consistency","Mamba","frame interpolation","cross-frame attention"],"falsifier":"Train Stage 2 on arbitrarily corrupted volumes rather than Stage 1 outputs, or test Stage 2 on a new motion pattern such as a different joint or a larger flexion range, and measure whether the PSNR and SSIM gains over the w/o SC ablation persist; if the gains drop to near zero when Stage 1's artifact distribution shifts, the spatial-consistency correction is memorization rather than general volumetric coherence.","tokens_in":1585,"feed_emoji":"🩻","tokens_out":1844,"duration_ms":63588,"temperature":0.7,"pith_summary":"TSSC-Net claims that a two-stage deep learning pipeline can turn just two MRI frames into a complete 12-frame 4D dynamic sequence, achieving 6x temporal super-resolution in a single inference step while correcting the cross-slice inconsistencies that arise when frames are generated slice by slice. The paper argues this matters because registration-based interpolation, the standard way to synthesize intermediate frames, fails on large-amplitude motion. On an in-house 12-case knee dataset with flexion from 0 to 80 degrees, TSSC-Net reports PSNR 20.041 dB and SSIM 0.730 versus 17.580 dB and 0.635 for the best registration-based baseline. On the public ACDC cardiac dataset the method is competitive, with PSNR 32.971 dB and SSIM 0.977, showing that the gains are largest when motion is fast and large.","feed_headline":"Two MRI frames yield a 12-frame 4D scan","feed_subtitle":"On fast-moving knees it beats registration-based interpolation by roughly 2.5 dB PSNR.","key_machinery":"The load-bearing object is the two-stage pipeline. Stage 1 is a transformer-based diffusion model whose reverse process is conditioned on the start frame $I_0$ and end frame $I_1$; cross-frame attention lets the denoiser use both boundary frames to synthesize intermediate frames, and DDIM acceleration is used at inference. Stage 2 is a spatial consistency enhancement network built from residual tri-directional Mamba blocks, each scanning the 3D volume along three orderings of the x, y, and z axes (SSM$_{xyz}$, SSM$_{yzx}$, SSM$_{zxy}$) to capture long-range dependencies in all directions. It is trained with a composite loss combining MSE, wavelet transform, and total variation regularization. The crucial work is that Stage 1 supplies temporally plausible but slice-wise inconsistent volumes, and Stage 2 acts as a learned 3D post-processor restoring volumetric coherence.","core_discovery":"The central claim is that a diffusion model conditioned on the start and end frames of a 4D MRI scan can generate all intermediate frames, providing 6x temporal super-resolution in a single inference step, and that a second network using residual tri-directional Mamba blocks can repair the cross-slice inconsistencies created by generating slices independently. The paper reports that on fast, large-motion knee MRI this beats the registration-based baselines by a wide margin, while on low-motion cardiac MRI it matches the best baseline. The ablation study without the spatial consistency network drops PSNR by about 3.6 dB on cardiac and 1.9 dB on knee, which the paper attributes to the spatial consistency network correcting cross-slice misalignment.","pith_inferences":["An implication the authors leave implicit is that the same architecture could be applied to other dynamic volumes with sparse temporal sampling, such as 4D CT or fetal MRI, as long as boundary frames bracket a coherent motion.","The 6x single-step interpolation claim suggests testing on larger temporal leaps, such as 8x or 12x, to see whether cross-frame attention degrades gracefully; the paper does not report such scaling.","If Stage 2 generalizes as a true consistency corrector, it could be attached to any slice-wise generation method, not just this diffusion Stage 1; this is a testable extension the paper does not make.","The reported knee gains come from only 12 scans, so the practical magnitude for clinical use depends on whether the gap holds on a larger multi-center cohort, which the paper itself leaves for future work."],"forward_implications":["A complete 12-frame 4D dynamic sequence can be produced from only two acquired frames, potentially reducing the scan-time burden of 4D MRI.","The method should be most clinically valuable where motion is fast and large, such as knee flexion or cardiac phases with large deformation, since the reported gains over registration-based baselines are largest there.","The two-stage design implies that temporal coherence and spatial coherence can be treated separately: diffusion handles the temporal gap, and tri-directional Mamba handles cross-slice consistency.","On low-motion data the method is competitive but not clearly superior, so the claimed advantage is specifically for large-motion scenarios.","Cross-slice errors from independently generated slices are correctable by a learned 3D refinement network, according to the ablation study."],"supporting_citations":[{"why":"VoxelMorph registration baseline that TSSC-Net must beat on the large-motion knee dataset.","marker":"[2]"},{"why":"TransMorph transformer-based registration baseline that achieves the second-best knee results.","marker":"[5]"},{"why":"UVI-Net unsupervised bidirectional interpolation baseline that TSSC-Net beats on knee but loses slightly to on cardiac PSNR.","marker":"[16]"},{"why":"Denoising diffusion probabilistic models, which supply the forward and reverse diffusion process that Stage 1 adapts.","marker":"[12]"},{"why":"Mamba state-space model architecture that the tri-directional blocks in Stage 2 are built on.","marker":"[10]"},{"why":"ACDC public cardiac MRI dataset used for the low-motion evaluation.","marker":"[4]"},{"why":"Supports the practice of conditioning the diffusion reverse process on known start and end frames for video generation.","marker":"[17]"}],"fun_headline_variants":["Diffusion model builds 6x faster 4D MRI from just 2 frames","Two MRI frames, one diffusion pass: 12-frame 4D scan","TSSC-Net: 6x temporal SR for 4D MRI with spatial fixes","Fast motion? Diffusion fills 4D MRI gaps, beats interpolation","One-shot 6x temporal super-resolution for 4D MRI"],"cache_read_input_tokens":9856,"weakest_assumption_plain":"The whole two-stage claim rests on Stage 2 learning to correct general cross-slice inconsistencies rather than memorizing the specific artifacts Stage 1 produced during training, but the paper never specifies how Stage 2's training pairs were constructed.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model builds 6x faster 4D MRI from just 2 frames","Two MRI frames, one diffusion pass: 12-frame 4D scan","TSSC-Net: 6x temporal SR for 4D MRI with spatial fixes","Fast motion? Diffusion fills 4D MRI gaps, beats interpolation","One-shot 6x temporal super-resolution for 4D MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1251,"prompt_tokens":902,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":247}},"tokens_in":518,"tokens_out":349,"duration_ms":3831,"temperature":1.0,"reasoning_tokens":247,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:46:47.390062+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Stage 2 on arbitrarily corrupted volumes rather than Stage 1 outputs, or test Stage 2 on a new motion pattern such as a different joint or a larger flexion range, and measure whether the PSNR and SSIM gains over the w/o SC ablation persist; if the gains drop to near zero when Stage 1's artifact distribution shifts, the spatial-consistency correction is memorization rather than general volumetric coherence.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"UVI-Net unsupervised bidirectional interpolation baseline that TSSC-Net beats on knee but loses slightly to on cardiac PSNR."}],"review_version":1}