{"id":"b534beec-bf2d-4ddd-9e73-a5f526587fda","arxiv_id":"2505.04380","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding a second decoder that fuses encoder and first-decoder features improves deformable image registration accuracy on LPBA40, IXI, and OASIS, with gains of 0.4 to 2.3 Dice points.","lead":"This paper proposes a medical image registration network with one encoder and two decoders, where the extra decoder refines the deformation field from the first. The authors report small but consistent Dice gains on three brain MRI benchmarks and across several popular U-Net-like baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pretraining confound in baselines may dominate the reported gains; a same-protocol comparison could settle it.","rationale":"The reader identified the same weakness: baselines are not described as pretrained while the proposed models are, and the 9-volume LPBA40 test set lacks variance estimates. My reading of the full text confirms this is the most load-bearing issue. Section 4.2.1's own ablation shows pretraining yields small but consistent gains (0.2–0.3% Dice), which is the same order of magnitude as some of the claimed improvements (e.g., U-UNet 0.671 vs. VoxelMorph 0.658; the +1.3% gap in Table 4). Since the architecture change also increases parameters (U-UNet 0.652M vs. VoxelMorph 0.359M), the gain could partly come from capacity and partly from the decoder design; the paper does not control for parameter count either. The internal inconsistency noted by the reader (VoxelMorph %|J|<0 differs between Table 4 and Table 5: 0.384 vs. 0.288) is real but secondary; it suggests possible different training runs or data splits for the same baseline, which further undermines the controlled comparison. The proposed test directly addresses the confound: retrain baselines with the same pretraining protocol. If the gain persists, the central claim holds; if it shrinks, the paper's main conclusion is overstated. Therefore the reader's CONDITIONAL verdict is appropriate; I do not see grounds for outright rejection, but the current evidence is insufficient for full acceptance.","tokens_in":14709,"tokens_out":1167,"duration_ms":11849,"concrete_test":"Run the baseline VoxelMorph and TransMorph with the identical two-stage pretraining protocol: first train encoder+Dec1 to convergence, then continue training the full baseline (with no Dec2) for the same number of epochs and with the same learning rate schedule as the Tetrahedron-Net variants. If the pretrained baselines close the gap to within 0.2% Dice on LPBA40 and IXI, the decoder's specific contribution is negligible; if the gap persists above ~0.5% Dice across three seeds, the architectural claim is supported. Also report the standard deviation over at least three random seeds for the 9-volume LPBA40 test set.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that appending a second decoder (Dec2) yields consistent, architecture-independent gains. The key comparisons in Table 4 and Table 5 compare the proposed Tetrahedron-Net variants against baselines (VoxelMorph, ViT-V-Net, TransMorph, TransMorph-bspl) that are presumably trained from scratch in a single stage. However, Section 4.2.1 explicitly states that the proposed models load a pre-trained encoder and first decoder (Dec1) before adding Dec2, and that this pretraining alone improves Dice by 0.2–0.3% on LPBA40 (Table 1: e.g., U-UNet from 0.665 to 0.667, U-UNet++ from 0.672 to 0.675). The main results in Table 4/Table 5 are reported for the pretrained versions, while the baselines are not described as receiving any equivalent warm-start. If pretraining alone accounts for part or all of the observed gap (e.g., VoxelMorph baseline 0.658 vs. U-UNet 0.671 in Table 4, a 1.3% gap; on IXI, Tetrahedron-Net(VM) 0.749 vs. VoxelMorph 0.729, a 2.0% gap), then the specific contribution of the two-level decoder architecture is overstated. The paper's own ablation (Section 4.2.4) says 'These models are loaded with pre-trained Enc and Dec1', but the baseline VoxelMorph in Table 4 is not listed as pretrained. No error bars or repeated-run variance are reported anywhere; on LPBA40 the test set is only 9 volumes, so the reported differences (e.g., 0.671 vs. 0.665) may be within run-to-run variability. This is the load-bearing weakness: the controlled comparison required to isolate the decoder's contribution (same pretraining, same training schedule, same number of parameters or at least reported variance) is missing, so the central 'consistent performance gains' claim is not yet established as specifically due to the Dec2 architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Tetrahedron-Net, an extension of U-Net-like architectures for unsupervised 3D medical image registration. The central idea is to append a second decoder that receives skip connections from both the encoder and the first decoder, enabling a coarse-to-fine prediction of the deformation field. The authors instantiate three versions of the second decoder (UNet++, UNet3+, DenseUNet) and also integrate the two-level decoder into VoxelMorph, ViT-V-Net, TransMorph, and TransMorph-bspl. Experiments are reported on LPBA40, IXI, and OASIS, showing consistent improvements in Dice score and reductions in the percentage of non-positive Jacobian determinants. Ablation studies examine the impact of pretraining, encoder skip connections, the number of decoder levels, and the structure of the second decoder.","tokens_in":15271,"tokens_out":4024,"duration_ms":39812,"significance":"If the reported gains hold under a controlled comparison, the contribution is valuable: the two-level decoder is a simple, architecture-agnostic modification that can be bolted onto any U-Net-like registration network, and the paper demonstrates it on four public backbones across three datasets. The ablations are a strength, as they attempt to isolate the contribution of the second decoder and its skip connections. However, the current evaluation does not establish that the gains come from the decoder itself rather than from the pretraining schedule, and the lack of any variance or significance reporting makes the small absolute improvements difficult to interpret. The work is therefore potentially significant but not yet convincingly supported.","major_comments":[{"comment":"The main comparison is confounded by pretraining. Section 4.2.4 states that the proposed models are 'loaded with pre-trained Enc and Dec1', and Table 1 shows that this pretraining alone improves Dice by 0.2–0.3% on LPBA40 (e.g., U-UNet from 0.665 to 0.667). The baselines in Tables 4 and 5 (VoxelMorph, ViT-V-Net, TransMorph) are not described as receiving any equivalent warm-start. Since the reported gains of Tetrahedron-Net over these baselines are 1.0–2.3% (e.g., VoxelMorph 0.658 vs. U-DenseUNet 0.681 in Table 4), part or all of the advantage could be due to the pretraining stage rather than the second decoder. The authors should rerun the baselines under the same two-stage pretraining protocol, or train all models from scratch, to isolate the architectural contribution.","section":"§4.2.1, §4.2.4, Tables 1, 4, 5"},{"comment":"No measures of uncertainty are reported anywhere. The LPBA40 test set has only 9 volumes, and the IXI and OASIS splits are also relatively small. The reported differences (e.g., 0.671 vs. 0.658 for U-UNet vs. VoxelMorph in Table 4) could easily fall within run-to-run variability. The paper should report mean and standard deviation over at least three independent training runs with different seeds, and ideally include paired statistical tests (e.g., Wilcoxon signed-rank) for the Dice comparisons. Without this, the central claim of 'consistent performance gains' is not statistically supported.","section":"Tables 4 and 5"},{"comment":"There is an internal inconsistency in the reported %|J|<0 for VoxelMorph on LPBA40: Table 4 lists 0.384, while Table 5 lists 0.288 for the same dataset and baseline. Since the paper uses the Jacobian metric to claim improved regularity of deformation fields, this discrepancy undermines the reliability of that claim. The authors should clarify which value is correct and ensure that all tables are consistent.","section":"Tables 4 and 5"},{"comment":"The choice of DenseUNet as the second decoder for the main results in Table 5 is made based on LPBA40 test-set performance (Section 4.2.4 states that DenseUNet 'achieves the best results'), and the default two-level decoder configuration is similarly selected based on LPBA40 (Section 4.2.3). This is a form of model selection on the test set, which can inflate results on LPBA40 and may not generalize. The authors should either select the Dec2 variant using a validation set or report results for all Dec2 variants on IXI and OASIS to demonstrate that the choice is not overfitted to LPBA40.","section":"§4.2.4, §4.3"}],"minor_comments":[{"comment":"The description of the UNet++-based second decoder says 'each decoder layer is connected to all preceding decoder layers', but the general equation in Section 3.2.3 (Eq. 5) only concatenates the immediately previous Dec1 and Dec2 features. The relation between the specific decoder structures and the general formulation should be clarified, either by giving separate equations for each variant or by stating that Eq. (5) is the simplest case.","section":"§3.3, §3.4"},{"comment":"The 'Level of decoder 1' entry reports DSC of 0.657, while the U-UNet baseline in Table 1 without pretraining is 0.665 and in Table 2 without encoder features is 0.667. The relationship between these numbers is not explained; if the ablation in Table 3 uses a different configuration (e.g., without pretraining or with different skip connections), this should be stated explicitly.","section":"Table 3"},{"comment":"Please state the number of random seeds used, the hardware/software versions, and the approximate training time. No code or trained models are provided, so these details are important for reproducibility. The current description ('one NVIDIA RTX 3090 GPU') is insufficient for an independent replication.","section":"§4.1.2"},{"comment":"The stated limitation ('validated only on medical image registration') is framed as a limitation, but the paper explicitly claims generality to other dense prediction tasks in the introduction; the more relevant limitation is the absence of a controlled ablation isolating the decoder's benefit from the pretraining schedule, as noted above. Consider rephrasing the limitations section to address the actual threats to validity.","section":"§5 'Limitations and future work'"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a straightforward architecture paper, and the core idea is simple enough to be of interest to the registration community. However, the evaluation protocol needs substantial revision: the pretraining confound and the absence of any variance statistics are load-bearing issues that preclude acceptance in their current form. The internal inconsistency in the Jacobian table is also concerning and should be corrected. I do not see grounds for rejection if the authors can provide a cleanly controlled comparison and uncertainty quantification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea is genuinely simple and worth taking seriously: append a second decoder that takes skip connections from both the encoder and the first decoder, and let it predict the deformation field in a coarse-to-fine way. Applied to VoxelMorph, ViT-V-Net, and TransMorph, the reported Dice gains are consistent in direction across LPBA40, IXI, and OASIS. That is a real result, and the paper deserves credit for running proper ablations: it checks the encoder skip connections, the number of decoders, and the pretraining step separately. The DenseUNet-style Dec2 variant consistently looks best, and the integration code appears straightforward to reproduce.\n\nThe soft spots are proportionate but not fatal. The load-bearing one is the pretraining confound. The proposed models are pretrained as Enc+Dec1 and then fine-tuned with Dec2, while the baselines are not described as receiving any warm-start. The paper's own Table 1 shows pretraining alone buys 0.2–0.3 Dice on LPBA40. Several reported gains over baselines are in the 0.4–2.3 range, so a same-protocol comparison (pretrain the baselines too, or train Tetrahedron from scratch) is needed before attributing the entire gap to the two-level decoder. The stress-test note gets this right. Second, no standard deviations or repeated runs are reported anywhere; on a 9-volume LPBA40 test set, a 0.2-point Dice difference could easily be noise. Third, there is an internal inconsistency for VoxelMorph on LPBA40: %|J|<0 is 0.384 in Table 4 but 0.288 in Table 5. Likely a typo, but it needs fixing. Finally, DenseUNet was selected as the default Dec2 based on LPBA40 test performance, then carried to IXI and OASIS; that is post-hoc model selection on the test set, mild but worth acknowledging.\n\nNone of this sinks the paper. The architecture is well motivated, the ablations are mostly clean, and the reported gains, while modest, are not implausible. What is missing is a controlled experiment plus error bars. The paper is worth a serious referee: the reviewer should ask for same-protocol pretraining comparisons, variance reporting, and correction of the Table 4/5 inconsistency. I would bring it to a reading group as an example of a simple architectural variant with a clear empirical claim, and I would cite it if I were writing a registration-backbone paper. Send it to peer review.","headline":"A simple, plausibly useful second-decoder plug-in for registration U-Nets, but the paper never isolates the decoder's contribution from a pretraining advantage and lacks variance reporting.","tokens_in":15715,"tokens_out":1679,"would_cite":true,"duration_ms":18072,"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":"The paper claims that appending a second decoder, which fuses encoder and first-decoder features, to any U-Net-like registration backbone yields consistent accuracy gains on LPBA40, IXI, and OASIS.","keywords":["medical image registration","deformable registration","U-Net","two-level decoder","skip connections","deformation field","unsupervised learning","brain MRI"],"falsifier":"Retrain each comparison model with the same two-stage pretraining schedule used for Tetrahedron-Net (encoder and first decoder first, then the full network) and compare Dice on LPBA40; if the gap essentially disappears, the second decoder is not the source of the gain.","tokens_in":14548,"feed_emoji":"🧠","tokens_out":11765,"duration_ms":88225,"temperature":0.7,"pith_summary":"This paper tries to establish that the standard single-encoder, single-decoder U-Net leaves registration accuracy on the table, and that the cheap fix is a second decoder rather than more elaborate skip connections. Tetrahedron-Net keeps one shared encoder and decodes twice: the first decoder behaves like a normal U-Net decoder, and the second decoder concatenates features from both the encoder and the first decoder at every scale. On three brain MRI benchmarks (LPBA40, IXI, OASIS), the paper reports higher Dice overlap and lower fractions of non-smooth deformation-field voxels than VoxelMorph, ViT-V-Net, and TransMorph. The paper's practical claim is that the two-level decoder is a drop-in upgrade that improves any U-Net-like registration backbone, which would matter because it converts an architecture redesign into a simple append step.","feed_headline":"One extra decoder improves every U-Net-style registration model","feed_subtitle":"Appending the same two-level decoder to VoxelMorph, ViT-V-Net, or TransMorph raises Dice on all three brain MRI sets.","key_machinery":"The load-bearing piece is the two-level decoder, a second decoder branch appended to an off-the-shelf encoder-decoder backbone. In the default U-UNet instantiation, Decoder 2 mirrors Decoder 1's structure but its $k$-th block takes as input the concatenation of the corresponding encoder feature, the previous Decoder-1 feature, and its own previous feature: $x^k_{\\mathrm{dec2}} = \\mathrm{Up}(\\mathrm{CR}([x^{L-k}_{\\mathrm{enc}}, x^{k-1}_{\\mathrm{dec1}}, x^{k-1}_{\\mathrm{dec2}}]))$. This turns registration into a coarse-to-fine refinement: Decoder 1 supplies a prior, and Decoder 2 fuses it with multi-scale encoder details to predict the final deformation field. The paper tests the same machinery with UNet++, UNet3+, and DenseUNet as second-decoder structures, reporting the best scores with the dense-block version.","core_discovery":"The paper's central claim is that a U-Net-style registration network produces better deformation fields when its decoder is replaced by two cooperative decoders sharing one encoder: the first decoder reconstructs multi-scale features exactly as in U-Net, and the second decoder refines that reconstruction by concatenating, at each scale, the encoder feature with the preceding outputs of both decoders. This makes the final prediction a coarse-to-fine refinement of the first decoder's result. The paper reports that using a DenseUNet-style second decoder gives the best configuration, reaching 0.681 average Dice on LPBA40 compared with 0.658 for the VoxelMorph baseline, and that the same appended-decoder recipe produces consistent gains when applied to VoxelMorph, ViT-V-Net, TransMorph, and TransMorph-bspl on LPBA40, IXI, and OASIS.","pith_inferences":["Editorial inference: because the appended decoder is agnostic to the encoder and loss function, the same recipe should transfer to other dense prediction tasks, such as medical image segmentation, which the authors list as future work.","Editorial inference: the reported comparisons do not isolate the decoder from the two-stage pretraining protocol, so a replication that pretrains single-decoder baselines identically is needed before the full improvement is attributed to Decoder 2.","Editorial inference: the LPBA40 evaluation uses one nine-volume test set with no error bars, so repeated runs with different seeds are needed to show whether the 0.6 to 2.3 point Dice differences are stable.","Editorial inference: the reported faster convergence with Decoder 2 suggests the second branch acts partly as an optimization aid, which could be separated from capacity effects by measuring epoch-wise accuracy in a low-data setting."],"forward_implications":["Because the appended decoder couples only through skip connections and the loss is unchanged, any U-Net-like registration network can adopt the two-level decoder with a minimal code change.","The reported gains hold for both convolutional (VoxelMorph) and transformer-based (ViT-V-Net, TransMorph) backbones, so the benefit does not depend on a particular encoder.","The internal design of the second decoder matters: a DenseUNet-style second decoder outperforms plain UNet, UNet++, and UNet3+ second decoders on LPBA40.","Adding more than two decoders continues to improve Dice but roughly doubles parameters per decoder, so the paper chooses two levels as the default trade-off.","Pretraining the encoder and first decoder before adding the second decoder adds a small but consistent gain across all second-decoder variants."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder template that Tetrahedron-Net extends with a second decoder.","marker":"Ronneberger, Fischer and Brox (2015)"},{"why":"Supplies the VoxelMorph registration framework and the convolutional baseline whose decoder is replaced.","marker":"Balakrishnan et al. (2019)"},{"why":"Supplies the UNet++ second-decoder variant with nested dense skip connections.","marker":"Zhou et al. (2019)"},{"why":"Supplies the UNet3+ second-decoder variant with full-scale feature fusion.","marker":"Huang et al. (2020)"},{"why":"Supplies the DenseUNet-style second decoder that gives the best reported scores.","marker":"Wu et al. (2021)"},{"why":"Supplies the ViT-V-Net transformer baseline upgraded by the appended decoder.","marker":"Chen et al. (2021a)"},{"why":"Supplies the TransMorph and TransMorph-bspl baselines upgraded by the appended decoder.","marker":"Chen et al. (2022)"},{"why":"Provides the spatial transformer network that warps the moving image with the predicted deformation field.","marker":"Jaderberg et al. (2015)"}],"fun_headline_variants":["Two decoders outshine one in U-Net medical image registration","Tetrahedron-Net: a second decoder refines U-Net registration","Add a decoder to U-Net and lift registration accuracy","Simple dual-decoder design improves all U-Net-style registrations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the reported gains to show what the paper claims, the comparison models must differ from Tetrahedron-Net only by the added decoder, but the proposed models were pretrained in two stages while the comparison models were not, so pretraining could be carrying part of the improvement.","fun_headline_variants_meta":{"raw":{"variants":["Two decoders outshine one in U-Net medical image registration","Tetrahedron-Net: a second decoder refines U-Net registration","Add a decoder to U-Net and lift registration accuracy","Simple dual-decoder design improves all U-Net-style registrations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000946,"raw_usage":{"total_tokens":4069,"prompt_tokens":1007,"completion_tokens":3062,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":2987}},"tokens_in":623,"tokens_out":3062,"duration_ms":22156,"temperature":1.0,"reasoning_tokens":2987,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:30:06.641863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain each comparison model with the same two-stage pretraining schedule used for Tetrahedron-Net (encoder and first decoder first, then the full network) and compare Dice on LPBA40; if the gap essentially disappears, the second decoder is not the source of the gain.","supporting_citations":[],"review_version":1}