{"id":"1562d2cd-8ac5-42e3-a6cc-76ff755df7e1","arxiv_id":"2411.15426","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LDM-Morph combines latent diffusion model features, a cross-attention module, and a hierarchical pixel-plus-latent loss to achieve higher Dice scores and fewer deformation folds in 2D cardiac image registration.","lead":"This paper introduces a deformable image registration method that uses a latent diffusion model to extract semantic features from medical images and a cross-attention module to combine them with global features, improving alignment accuracy and reducing folding. The method outperformed several existing registration algorithms on four public 2D cardiac datasets, with code released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 1–2% DSC gains over the best Transformer baselines are within inter-subject noise on 3 of 4 datasets; without significance testing the 'outperformed' claim is not supported.","rationale":"Reading the paper in good faith, the architecture is plausible, the ablation study is reasonable, and the code is released. The central claim, however, is an 'outperformed' statement. For it to hold, the accuracy differences must be real and not attributable to sampling noise. The paper's own Table 1 shows mean DSC differences over the best Transformer baselines of only 0.008–0.013, with standard deviations around 0.05–0.08 and test sizes of 50–1276. Approximate significance calculations from the published mean/SD and test sizes show that the differences are not significant on CAMUS-2CH (p≈0.16), CAMUS-4CH (p≈0.10), and ACDC (p≈0.54); only ECHO (large test) reaches significance. No statistical tests are reported, so the accuracy superiority claim overreaches. The reader's identified weakest assumption (unknown LDM pretraining data) is valid and worth investigating, but it is speculative; the significance issue is immediately visible from the paper's own numbers. Missing comparisons to recent SOTA methods such as H-ViT and FSDiffReg are also relevant but secondary to whether the reported gains are even statistically meaningful. Thus the most load-bearing concern is the lack of demonstrated statistical significance for the headline accuracy gain. This does not invalidate the method, but it should temper the conclusion to a more modest claim unless significance tests are provided. The reader's CONDITIONAL verdict remains appropriate, so no change to the verdict is recommended.","tokens_in":31441,"tokens_out":6930,"duration_ms":60706,"concrete_test":"Compute per-case DSC for LDM-Morph and for the best-performing Transformer baseline on each dataset (per-case values are summarized by mean and std in Table 1). Run paired permutation tests or Wilcoxon signed-rank tests on the per-case DSC vectors. If the null hypothesis of equal accuracy is rejected at α=0.05 on fewer than three datasets, the accuracy component of the 'outperformed' claim is not supported; the manuscript should be revised to claim competitive accuracy with improved topology preservation, or the required statistical evidence should be added.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LDM-Morph 'outperformed' TransMorph/TransMatch in registration accuracy rests on mean DSC differences of 0.8–1.3 percentage points (Table 1: 0.882 vs 0.871 on CAMUS-2CH; 0.889 vs 0.876 on CAMUS-4CH; 0.850 vs 0.842 on ACDC; only ECHO is 0.883 vs 0.874). Given the reported standard deviations (≈0.05–0.08) and test sizes (80, 80, 50, 1276), approximate paired tests give p≈0.16, 0.10, and 0.54 for the first three datasets, while ECHO is significant (p<0.001). The paper reports no significance tests, yet Section 5.1 and the abstract assert superiority in both accuracy and topology preservation. The topology-preservation component (|Jφ|≤0) shows larger relative differences (e.g., 0.239% vs 0.696% on CAMUS-2CH) and is more convincing, but the accuracy half of the claim is not established by the reported statistics. This is load-bearing because the claimed advantage over the strongest Transformer baselines is the principal differentiator of the proposed architecture; if the accuracy gains are within noise, the method reduces to a more complex architecture with comparable accuracy but better folding metrics, which is a weaker claim than 'outperformed'.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LDM-Morph, an unsupervised deformable registration method that combines features from a frozen latent diffusion model (LDM) with a Swin-style global feature stream through a latent-global cross-attention (LGCA) module, and trains with a hierarchical similarity loss evaluated in both pixel space and LDM latent space. The method is evaluated on four public 2D cardiac datasets (CAMUS-2CH, CAMUS-4CH, EchoNet-Dynamic, ACDC) against two classical algorithms and five deep-learning baselines, reporting Dice similarity coefficients, percentages of non-positive Jacobian determinants, and runtime. The authors claim state-of-the-art accuracy and topology preservation with comparable computational cost, and they include an ablation study plus a cross-dataset generalization experiment.","tokens_in":31788,"tokens_out":8328,"duration_ms":72270,"significance":"If the claims are substantiated, the paper would make a useful contribution: it shows that semantic features from a pretrained LDM can be injected into a registration network and used in a dual-space loss, with public code and experiments across multiple cardiac datasets. The strengths are the breadth of the evaluation, the inclusion of topology metrics, the cross-dataset generalization test, and the ablation isolating the three proposed components. However, the central 'outperformed' claim currently rests on mean DSC differences of about 1 percentage point over the strongest Transformer baselines, and the paper lacks significance testing; the method description also contains a concrete loss-formula inconsistency and leaves the LDM pretraining data unspecified. These issues are fixable, but they are load-bearing for the paper's headline claims.","major_comments":[{"comment":"The abstract and Section 5.1 claim that LDM-Morph 'outperformed' the Transformer-based baselines in registration accuracy, but the reported mean DSC advantages over TransMorph/TransMatch are 0.011 on CAMUS-2CH (0.882 vs 0.871), 0.013 on CAMUS-4CH (0.889 vs 0.876), and 0.008 on ACDC (0.850 vs 0.842), with standard deviations around 0.05–0.08 and test sizes of 80, 80, and 50. No significance tests are reported, and on these three datasets the gaps are of the same order as the sampling noise; only the ECHO comparison (0.883 vs 0.874, n=1,276) is clearly beyond noise. Please provide paired significance tests, with multiple-comparison correction, for both DSC and the |Jφ|≤0 metric, or soften the accuracy claim to a 'comparable accuracy with improved topology preservation' claim. As it stands, the accuracy half of the headline claim is not established by the reported statistics.","section":"Sec. 5.1, Table 1"},{"comment":"The balance weight β in Eq. (13) is selected using the test sets. Figure 7 plots DSC and |Jφ|≤0 on the four testing datasets for β from 0.2 to 1.0, and Section 4.2 then fixes β=0.6 'for all the experiments.' This means the final test-set numbers in Table 1 are the product of test-set hyperparameter tuning, which biases the comparison in favor of LDM-Morph. Select β on the validation folds (or via nested cross-validation), report the chosen value separately for each dataset, and freeze it before evaluating on the test sets. The observation that CAMUS-4CH has its best DSC at β=0.4 while 0.6 is used globally further highlights the need for a principled selection procedure.","section":"Sec. 5.3, Fig. 7 and Sec. 4.2"},{"comment":"The pretraining data for the LDM are never specified. Section 3.2 states that 'an LDM was pretrained based on the image pairs as described below,' and Section 4.2 only gives the architecture (VQGAN plus a U-Net with channel size 64). If the LDM was trained on the same four datasets, including their test folds, then the hierarchical loss and the cross-dataset generalization experiment in Section 5.2 are not independent, and the reported gains could partly reflect the LDM having memorized the test distribution. Please disclose the exact pretraining dataset(s), the train/validation/test split used for the LDM, and confirm that the LDM training data do not overlap the test sets used in Tables 1 and 2.","section":"Secs. 3.2 and 4.2"},{"comment":"The hierarchical similarity loss is double-weighted as written. Eq. (14) defines L_org as (1/|Ω|) Σ_p β [M∘φ^{-1}(p)−F(p)]^2, and Eq. (15) defines L_lat as (1/|Ω|) Σ_p (1−β) [E(M∘φ^{-1}(p))−E(F(p))]^2, while Eq. (13) then forms β L_org + (1−β) L_lat. The result is β^2 times the pixel MSE plus (1−β)^2 times the latent MSE, not the stated convex combination. This changes the interpretation of β in Figure 7 and Section 4.2. Please remove the inner β and (1−β) factors (or redefine L_org and L_lat as unweighted means), and state explicitly which form was actually implemented in the experiments.","section":"Sec. 3.4, Eqs. (13)–(15)"}],"minor_comments":[{"comment":"The posterior mean formula is inconsistent with the notation: α_t is defined in Eq. (2) as the cumulative product ∏_{s=1}^t (1−β_s), whereas the standard DDPM mean uses 1/√α_t with α_t = 1−β_t. Please correct the formula or align the notation with the standard DDPM expression.","section":"Sec. 3.2, Eq. (5)"},{"comment":"The smoothness term is an ordinary L2 gradient penalty on the deformation field; describing it as a 'diffusion-based regularizer' is misleading because no diffusion process is involved in that term.","section":"Sec. 3.4, Eq. (16)"},{"comment":"The feature-set notation '{ f i M L fF i|i = 1, 2,..., n }' is malformed and should be written as {f^i_{M_L}, f^i_F}; in the same passage, 'Marhov Chain' should be 'Markov chain.'","section":"Sec. 3.2"},{"comment":"The ablation rows are identified only by check marks, leaving the first row's configuration (the baseline without any of the three modules) implicit. Add explicit configuration labels to each row.","section":"Table 3"},{"comment":"The statement that LDM-Morph achieved the 'lowest standard deviation in DSC values' is true from Table 1, but the differences in standard deviations are very small (0.001–0.003); please avoid overinterpreting these differences without a formal variance comparison.","section":"Sec. 5.1"},{"comment":"In the sentence defining self-attention, 'matrics' should be 'matrices.'","section":"Eq. (8) and surrounding text"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a medical image analysis venue and the proposed approach is plausible, but the revision must address the significance-testing gap, the test-set selection of β, the unspecified LDM pretraining data, and the double-weighted loss formula. These are all addressable within the manuscript's scope, so I do not recommend rejection; however, the claims in the abstract and Section 5.1 need to be brought in line with the statistical evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading. It combines a frozen LDM feature extractor with a global cross-attention module and a pixel+latent loss, and tests it on four cardiac datasets with ablations and cross-dataset transfer. The code is public. That is a respectable package.\n\nWhat's genuinely new is the specific combination, and the ablation shows each piece adds something. The topology-preservation result (lower |J|≤0 percentages) is the most convincing part: the differences there are large and consistent.\n\nThe soft spots are real. First, the headline DSC gains over TransMorph/TransMatch are within inter-subject noise on CAMUS-2CH, CAMUS-4CH, and ACDC. With the reported SDs and test sizes, approximate paired tests give p≈0.16, 0.10, and 0.54; only ECHO is significant. The paper reports no significance tests, so the abstract's 'outperformed in accuracy' is not supported as written. Second, β is chosen after looking at test-set curves (Fig. 7, Sec. 5.3), which is a form of test-set tuning. Third, Eqs. (14) and (15) contain extra β and (1−β) factors, so the effective loss weights are squared; that is likely a typo but should be fixed. Fourth, the paper does not say what data the LDM was pretrained on. If the LDM saw the same cardiac datasets, the cross-dataset generalization experiment is not fully independent. Finally, the closest prior work FSDiffReg is cited but not compared; that is an omission worth addressing.\n\nNone of these is fatal. The architecture is coherent, the experiments are more extensive than most registration papers, and the limitations section is honest. With paired significance tests, a validation-based β selection, and a clarification of the LDM pretraining data, the accuracy claim could be put on solid ground. I would treat the folding improvement as the more robust contribution.\n\nWho should read it: anyone working on diffusion-guided registration or on 2D cardiac alignment. It is a useful data point even if the accuracy gains are partly noise. Send it to peer review; a good reviewer can push for the fixes.","headline":"Solid LDM-based registration recipe, but the accuracy claim over Transformers is statistically shaky; the folding gains are the stronger result.","tokens_in":32283,"tokens_out":3143,"would_cite":true,"duration_ms":27716,"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":"LDM-Morph claims that injecting latent diffusion features into a dual-stream encoder with cross-attention and a pixel-plus-latent similarity loss outperforms existing CNN- and Transformer-based deformable registration methods on four…","keywords":["deformable image registration","latent diffusion model","cross-attention","unsupervised learning","cardiac imaging","topology preservation","hierarchical similarity loss","Swin Transformer"],"falsifier":"A reader could retrain the LDM on only the training splits of the four datasets, or use an LDM pre-trained on a completely unrelated image collection, and then rerun the registration experiments; if the Dice advantage over TransMatch disappears or the folding rate rises, the central claim is falsified.","tokens_in":31261,"feed_emoji":"🫀","tokens_out":2905,"duration_ms":26513,"temperature":0.7,"pith_summary":"The paper tries to establish that deformable image registration improves when the network is given high-level semantic features from a pre-trained latent diffusion model, not just pixel-level features. To do this, it proposes LDM-Morph, an unsupervised framework with three components: an LDM-based feature extractor, a latent-and-global cross-attention module that fuses LDM features with Swin-style global features, and a hierarchical similarity loss evaluated in both pixel and latent space. The claim is that this combination yields higher Dice similarity coefficients and fewer deformation folds than the compared CNN-, Transformer-, and diffusion-based registration methods on four 2D cardiac datasets. If true, the paper would show that semantic information from generative models can be a practical ingredient in registration, improving both accuracy and topology preservation.","feed_headline":"Latent diffusion features beat CNN and Transformer registration","feed_subtitle":"A frozen latent diffusion model as feature extractor raises Dice overlap and reduces deformation folding on cardiac imaging.","key_machinery":"The key machinery is the LDM-based latent feature extraction module, which uses a frozen pre-trained latent diffusion model to encode moving and fixed images into latent variables via DDIM inversion and then extracts features from four convolutional layers of the denoising U-Net. These latent features are fed into the upper stream of a dual-stream encoder, while a lower stream computes Swin-style windowed self-attention; the LGCA module then performs window-based multi-head cross-attention by swapping queries between the two streams. The loss combines pixel-space MSE with latent-space MSE computed through the LDM encoder, weighted by beta = 0.6, plus a gradient regularizer.","core_discovery":"The central claim is that LDM-Morph outperformed existing state-of-the-art CNN- and Transformer-based registration methods in accuracy and topology preservation on the CAMUS-2CH, CAMUS-4CH, EchoNet-Dynamic, and ACDC datasets. The paper reports average Dice values of 0.882, 0.889, 0.883, and 0.850 respectively, with lower percentages of non-positive Jacobian determinants than most learned baselines, and better cross-dataset generalization when trained on EchoNet and tested on CAMUS. The authors attribute this to the LDM features enriching semantic information, the LGCA module enhancing interaction between latent and global features, and the latent-space similarity term reducing folding.","pith_inferences":["The paper does not specify which images were used to pre-train the LDM; if the same cardiac datasets were used, including test pairs, the reported gains may partly reflect memorization rather than generalizable semantic features.","The hierarchical loss behaves like a perceptual regularizer, so a similar benefit might appear in other modalities with low contrast or high noise, such as ultrasound or low-dose CT.","The method could likely be made end-to-end trainable by fine-tuning the LDM encoder, which the authors list as a limitation; this might yield larger gains but would also change the unsupervised setup.","A direct 3D extension is nontrivial because the LDM features are extracted from a 2D architecture, but applying the same idea with a 3D autoencoder is a plausible testable direction."],"forward_implications":["Using semantic features from a pre-trained latent diffusion model can improve deformable registration accuracy over CNN- and Transformer-only encoders.","Adding a latent-space similarity term to the loss reduces deformation folding, as measured by non-positive Jacobian determinants.","The framework generalizes across cardiac datasets when trained on one dataset and tested on another.","The improvements are incremental but consistent: roughly 1–3% higher Dice than Transformer baselines and larger margins over CNN and diffusion baselines.","The ablation study suggests that each added component—LDM features, LGCA, and the latent loss—contributes a measurable gain."],"supporting_citations":[{"why":"Defines the latent diffusion model architecture whose encoder and denoising U-Net are used for feature extraction.","marker":"Rombach et al., 2022"},{"why":"Provides the Swin Transformer windowed self-attention mechanism used in the global feature extraction module.","marker":"Liu et al., 2021"},{"why":"Supplies VoxelMorph, a CNN baseline that LDM-Morph is compared against.","marker":"Balakrishnan et al., 2019"},{"why":"Supplies TransMorph, a Transformer baseline that LDM-Morph is compared against.","marker":"Chen et al., 2022"},{"why":"Supplies TransMatch, a multilevel dual-stream Transformer baseline and a direct comparator.","marker":"Chen et al., 2023"},{"why":"Supplies DiffuseMorph, a diffusion-model-based registration baseline.","marker":"Kim et al., 2022"},{"why":"Provides the VQGAN used as the LDM encoder for latent-space encoding and reconstruction.","marker":"Esser et al., 2021"},{"why":"Supplies SyN, a conventional optimization-based registration baseline.","marker":"Avants et al., 2008"},{"why":"Supplies LDDMM, a conventional diffeomorphic registration baseline.","marker":"Beg et al., 2005"}],"fun_headline_variants":["LDM-Morph: diffusion features sharpen medical image alignment","Semantic diffusion features boost registration accuracy and topology","Latent diffusion encoder lifts cardiac registration performance","Diffusion-guided registration beats CNN and Transformer baselines","Frozen LDM features improve deformable registration fidelity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pre-trained LDM encoder provides meaningful and generalizable semantic features for cardiac images, which is only true if the LDM was not effectively trained on the same images used for evaluation.","fun_headline_variants_meta":{"raw":{"variants":["LDM-Morph: diffusion features sharpen medical image alignment","Semantic diffusion features boost registration accuracy and topology","Latent diffusion encoder lifts cardiac registration performance","Diffusion-guided registration beats CNN and Transformer baselines","Frozen LDM features improve deformable registration fidelity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1236,"prompt_tokens":936,"completion_tokens":300,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":226}},"tokens_in":552,"tokens_out":300,"duration_ms":3287,"temperature":1.0,"reasoning_tokens":226,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:18:01.491597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could retrain the LDM on only the training splits of the four datasets, or use an LDM pre-trained on a completely unrelated image collection, and then rerun the registration experiments; if the Dice advantage over TransMatch disappears or the folding rate rises, the central claim is falsified.","supporting_citations":[{"cited_title":", author Zhao, A","cited_arxiv_id":null,"evidence_quote":"Supplies VoxelMorph, a CNN baseline that LDM-Morph is compared against."},{"cited_title":", author Frey, E.C","cited_arxiv_id":null,"evidence_quote":"Supplies TransMorph, a Transformer baseline that LDM-Morph is compared against."},{"cited_title":", author Zheng, Y","cited_arxiv_id":null,"evidence_quote":"Supplies TransMatch, a multilevel dual-stream Transformer baseline and a direct comparator."},{"cited_title":", author Han, I","cited_arxiv_id":null,"evidence_quote":"Supplies DiffuseMorph, a diffusion-model-based registration baseline."},{"cited_title":", author Rombach, R","cited_arxiv_id":null,"evidence_quote":"Provides the VQGAN used as the LDM encoder for latent-space encoding and reconstruction."},{"cited_title":", author Epstein, C.L","cited_arxiv_id":null,"evidence_quote":"Supplies SyN, a conventional optimization-based registration baseline."},{"cited_title":", author Miller, M.I","cited_arxiv_id":null,"evidence_quote":"Supplies LDDMM, a conventional diffeomorphic registration baseline."}],"review_version":1}