{"id":"25478290-f9e3-4212-86f2-697cf611df1e","arxiv_id":"2505.04938","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A purely convolutional pyramid network with residual feature and deformation-field fusion achieves state-of-the-art Dice scores on LPBA and OASIS brain MRI registration benchmarks.","lead":"FF-PNet is a new brain image registration network that pairs two fusion modules, one for features and one for deformation fields, inside a three-layer pyramid. On two public brain MRI datasets it reports slightly higher Dice overlap than current methods while using a pure convolutional encoder.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on cross-paper baselines with an inconsistent ROI label count; a label-set or protocol mismatch could erase the 0.5–0.7 point DSC edge.","rationale":"The reader's weakest assumption identified the core issue: quoted baseline numbers without a demonstrated unified protocol. My stress-test agrees and adds a concrete internal inconsistency that makes the concern more specific: Section IV-A states LPBA has 54 ROIs while Table I is labeled 56 ROIs, and the reported winning margins are small relative to typical protocol-induced variation. I also note that FF-PNet does not dominate on SDlogJ or HD95, weakening the abstract's 'consistently outperforms' phrasing. I found no fundamental flaw in the architecture, and the ablations are internally coherent. The right response is not rejection but a conditional verdict requiring controlled baseline re-runs or explicit protocol documentation before the comparative claim is accepted.","tokens_in":10196,"tokens_out":3718,"duration_ms":42120,"concrete_test":"Re-run the nearest competitors (RDN, DualPRNet++, PIVIT, and H-ViT) using their released code on the exact LPBA and OASIS preprocessed scans, affine-initialized volumes, train/test splits, and ROI label sets used to produce FF-PNet's Tables I and II. If, under that unified protocol, FF-PNet's Dice advantage over every baseline shrinks below one standard deviation or reverses, the central superiority claim is not established; if the gap persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FF-PNet's Dice scores in Tables I and II establish a better accuracy-to-complexity trade-off than prior pyramid and transformer methods. The load-bearing condition is that these comparisons are apples-to-apples: identical preprocessing, affine initialization, train/test splits, and ROI label definitions. That condition is not demonstrated. Tables I and II mix in-house results with numbers quoted from the original papers, and no code or unified benchmark protocol is released. The manuscript itself flags a label-set inconsistency: Section IV-A describes LPBA as containing 54 manually labeled ROIs, while Table I is headed 'LPBA (56 ROIs)'. If the 56-ROI label set differs from the one used by RDN, PIVIT, or DualPRNet++, the reported Dice numbers are not comparable. The margins are also small: on LPBA, FF-PNet is 0.726 vs RDN's 0.720, and on OASIS, 0.884 vs DualPRNet++'s 0.879. These gaps are comparable to typical cross-protocol variation in brain registration, and the paper's own tables show FF-PNet does not win on every metric (RDN has lower HD95 on LPBA; PIVIT has lower SDlogJ on both datasets). The abstract's claim that the network 'consistently outperforms popular methods in metrics like Dice' is therefore not yet supported by controlled evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FF-PNet is an unsupervised deformable image registration network with a three-level pyramidal CNN encoder and a dual-stream decoder composed of a Residual Feature Fusion Module (RFFM) and a Residual Deformation Field Fusion Module (RDFFM). The encoder is purely convolutional, with no attention or MLP modules. The decoder predicts deformation fields at multiple scales, combining feature-level and field-level residual flows. The network is trained with an NCC similarity loss plus a diffusion regularizer. On the LPBA40 and OASIS datasets, the authors report mean Dice scores of 0.726 and 0.884, which are higher than the previously published numbers for VoxelMorph, TransMorph, DualPRNet++, Im2Grid, PIVIT, RDN, and H-ViT listed in their tables. Ablations are provided for the two fusion modules, the number of layers, and the initial channel count.","tokens_in":10470,"tokens_out":8953,"duration_ms":78190,"significance":"The proposed architecture is attractive because it achieves competitive accuracy with a simple convolutional design, and the ablations demonstrate that both RFFM and RDFFM contribute to the reported Dice gain (Table III). The layer and channel ablations (Tables IV and V) provide useful information about the complexity-accuracy trade-off within the proposed model. If the improvements over the baselines are confirmed in a controlled comparison, the dual-stream fusion decoder would be a valuable contribution to the registration toolbox, especially for settings where transformer-based models are too heavy. However, the central empirical claim currently depends on numerical results quoted from other papers without a demonstrated common evaluation protocol, and the margins are small; the paper's own tables show that FF-PNet does not win on every metric. Consequently, the significance is real but conditional on the evaluation being made rigorous.","major_comments":[{"comment":"The comparison against prior work is not demonstrated to be apples-to-apples. The tables mix FF-PNet results obtained in-house with numbers taken from the original publications, and the manuscript does not state that preprocessing, affine initialization, training/test splits, or ROI label definitions are identical. Section IV-A states that LPBA contains '54 manually labeled regions', while Table I is headed 'LPBA (56 ROIs)'; this inconsistency raises the possibility that the label set used for FF-PNet differs from that used by some baselines. Because the reported advantage over the best baseline is only 0.6 DSC points on LPBA (0.726 vs 0.720) and 0.5 points on OASIS (0.884 vs 0.879), a label or protocol mismatch could erase the improvement. The paper should re-run the baselines under a unified protocol, or provide a clear justification (e.g., a reference to identical preprocessing and label definitions) showing that the quoted numbers are directly comparable.","section":"§IV-C, Tables I and II"},{"comment":"No statistical significance tests are reported. The differences between FF-PNet and the strongest baselines are small relative to the reported standard deviations (e.g., on LPBA, RDN 0.720±0.011 vs FF-PNet 0.726±0.010; on OASIS, DualPRNet++ 0.879±0.015 vs FF-PNet 0.884±0.015). Since the test sets contain only 90 pairs (LPBA) and 38 pairs (OASIS), a paired significance test (e.g., Wilcoxon signed-rank) or at least a report of the per-subject performance distribution is needed to support the claim that the network 'consistently outperforms' popular methods. Without this, the reported improvements could be sampling noise.","section":"§IV-C, Tables I and II; abstract"},{"comment":"The abstract's claim that the method 'consistently outperforms popular methods in metrics like the Dice Similarity Coefficient' is not supported by the paper's own tables. On LPBA, RDN achieves a lower HD95 (3.756 mm vs 3.767 mm), and on both datasets PIVIT achieves a lower SDlogJ (0.246 vs 0.297 on LPBA; 0.414 vs 0.466 on OASIS). The text should be qualified to state that FF-PNet attains the highest Dice on these two datasets, with some other metrics still favoring earlier methods.","section":"Abstract; §IV-C"},{"comment":"The multi-scale deformation field composition is not fully specified. Eq. (4) computes F^{2'}_w = Warp(F^{2'}_m, ϕ3) and ϕ2 = Warp(ϕ3, ϕ') + ϕ', where ϕ3 is at the coarsest scale while F^{2'}_m and ϕ' are at a finer scale. The text does not describe how ϕ3 is resampled or upsampled to the finer grid before the warping and addition operations, nor how the 'superposition' of the fields mentioned in §III-A is computed. This omission makes the core contribution of RDFFM irreproducible from the description; please define the resampling method and the exact operation used to combine the fields.","section":"§III-C, Eq. (4)"}],"minor_comments":[{"comment":"The number of ROIs in LPBA is inconsistently reported as 54 in the text and 56 in Table I; please unify and justify this number.","section":"§IV-A vs Table I"},{"comment":"For OASIS, the table header says '35 ROIs' but the text does not mention the ROI count; please state it explicitly and cite the label atlas used.","section":"§IV-A"},{"comment":"The Recursive Decomposition Network (RDN) is cited as [24] in the text but as [27] in Table I; reference [24] in the bibliography is a different paper (Lv et al.). Please correct the citation.","section":"§II-B, related work"},{"comment":"There is a missing closing parenthesis in the third line of Eq. (3): 'Warp(ψ3, fθ(ψ')' should be 'Warp(ψ3, fθ(ψ'))'.","section":"Eq. (3)"},{"comment":"The sentence 'Experimental results show that when the network adopts a four-layer architecture, the image distorted by the deformation field of the penultimate layer has the highest similarity...' refers to an experiment not described elsewhere in the paper; it would be helpful to cite the corresponding table or explain the observation.","section":"§III-A"},{"comment":"The abbreviation 'CnC' is used before it is defined; please define it (presumably concatenation followed by convolution) at its first occurrence.","section":"§III-B and Figure 3"},{"comment":"The box plots in Figures 6 and 7 are not referenced in the running text; add a sentence or two explaining the regions shown and the observations to be drawn from them.","section":"Figures 6 and 7"},{"comment":"In the text after Table III, the claim that the full model 'achieves superior performance across all evaluation metrics' is stronger than the table shows, since Table III reports only DSC and HD95, not SDlogJ.","section":"§IV-D, Table III"}],"recommendation":"major_revision","confidential_remarks":"The paper's main architectural idea is plausible and the ablations are informative, but the evaluation needs to be made rigorous before the SOTA claim can be accepted. The label-count inconsistency (54 vs 56 ROIs) is a concrete warning sign that the cross-paper comparisons may not be protocol-matched. I would ask the authors to either rerun all baselines under their own unified protocol or provide evidence that the quoted numbers are directly comparable; otherwise the contribution should be framed as a relative improvement under a specific protocol. In addition, the promise that code will be released only 'if the paper is accepted' is not adequate for reproducibility; the editor may want to make code release a condition of acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: FF-PNet is a sensible pure-CNN pyramid registration architecture, and the ablation study supports the value of both decoder modules. The headline SOTA claim, though, rests on quoted baselines with no evidence of a common protocol, and the paper contradicts itself on LPBA ROI count (54 vs 56). So: solid architecture paper, unproven SOTA claim.\n\nWhat's new: the specific two-stream decoder with RFFM and RDFFM is not in the prior work they cite. RFFM warps features to create residual and cross-fused streams; RDFFM does two-stage warping and fusion on deformation fields. Both use only ordinary convolutions, no attention or MLP, which is attractive for deployment. The ablations are the strongest part: removing either module drops DSC on LPBA by 2–5 points, and the channel sweep justifies the 48-channel choice. The results are internally consistent.\n\nSoft spots, in rough order: (1) Tables I and II mix in-house runs with numbers quoted from RDN, PIVIT, DualPRNet++, and others. No code, no unified preprocessing, no significance tests. The margins over the best baselines are 0.6–0.7 Dice points, inside normal cross-paper variation. (2) Section IV-A says LPBA has 54 ROIs, while Table I is headed 56 ROIs. If the label set differs from what prior methods used, the comparison is apples to oranges. (3) The abstract claims the network \"consistently outperforms popular methods in metrics like Dice,\" but their own tables show PIVIT has lower SDlogJ on both datasets and RDN lower HD95 on LPBA. That overstates the case. (4) The paragraph about four-layer behavior at the start of Methods is confusing and not clearly connected to the layer ablation in Table IV. (5) The claim that the modules can be \"readily integrated into any pyramid-based registration framework\" is asserted without a plug-in experiment. These are real weaknesses, but none of them sink the architecture idea.\n\nWho this is for: the brain registration community and anyone designing lightweight decoders for registration networks. It deserves a serious referee: a good one can push for code, a rerun of baselines under a common protocol, and an explanation of the ROI discrepancy. I'd cite it as a clean pure-CNN pyramid reference, but not for its reported SOTA numbers.","headline":"Pure-CNN pyramid registration architecture with strong ablations, but SOTA Dice claim undercut by cross-paper baselines and a 54/56 ROI mismatch.","tokens_in":11004,"tokens_out":3283,"would_cite":true,"duration_ms":28244,"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":"FF-PNet shows a purely convolutional pyramid network with two residual fusion modules can outperform transformer-based registration methods on brain MRI benchmarks.","keywords":["deformable image registration","pyramid network","residual feature fusion module","residual deformation field fusion module","brain MRI registration","unsupervised registration","convolutional encoder","Dice similarity coefficient"],"falsifier":"Run VoxelMorph, TransMorph, and DualPRNet++ through FF-PNet's own training code, data splits, and hyperparameters on LPBA40 and OASIS, then compare Dice. If the re-run baselines match or exceed FF-PNet's 0.726 and 0.884 within one standard deviation, the claim of consistent improvement is not supported.","tokens_in":9996,"feed_emoji":"🧠","tokens_out":7898,"duration_ms":61685,"temperature":0.7,"pith_summary":"The paper argues that the accuracy bottleneck in deformable brain image registration lies not in the encoder but in how features and deformation fields are fused during decoding. It proposes FF-PNet, a three-layer pyramid network with a plain CNN encoder and two new decoder modules: RFFM, which cross-fuses warped and residual features, and RDFFM, which iteratively refines the deformation field. On LPBA40 and OASIS, the system reaches Dice scores of 0.726 and 0.884, above the reported numbers for VoxelMorph, TransMorph, DualPRNet++, PIVIT, Im2Grid, RDN, and H-ViT, with an 8.9-million-parameter model and no attention or MLP blocks. A sympathetic reading is that if this holds, lightweight convolutional pyramid designs with the right fusion blocks can match or beat much heavier transformer architectures for structural brain alignment.","feed_headline":"CNN pyramid with dual fusion beats transformer registration baselines","feed_subtitle":"Two plug-in decoder modules let a plain convolutional encoder rival attention models on LPBA and OASIS brain scans.","key_machinery":"Two residual fusion modules operating in parallel in the decoder. RFFM (Residual Feature Fusion Module) warps the moving feature using a coarse deformation field, combines it with the fixed feature, then warps the coarse fused feature again to form a residual feature, and cross-fuses the three streams through the Cross-Fusion Module (CFM). RDFFM (Residual Deformation Field Fusion Module) warps the deformation field itself, predicts an intermediate field, and adds it after another warp, making the field sensitive to local deformation details. The parallel operation lets one stream refine semantic correspondence while the other sharpens geometric detail.","core_discovery":"The central discovery is that parallel dual-stream fusion in the decoder is sufficient to close the accuracy gap with transformer-based registration. RFFM extracts residual and warped features from intermediate representations and cross-fuses them with the fixed image's features, while RDFFM treats the deformation field itself as a refined flow, warping it and adding an intermediate field. Over three pyramid scales, the network achieves the best reported Dice on both datasets among the listed methods, and ablations show each module adds roughly 1.3 to 1.8 points of Dice, with the largest gain from RFFM. The three-layer configuration with fusion outperforms a four-layer version at about one quarter of the parameters, and channel studies indicate that 48 initial channels maximize the accuracy-complexity trade-off.","pith_inferences":["A natural next test is to transplant RFFM and RDFFM into VoxelMorph's plain U-Net decoder and measure whether the Dice gain transfers; if it does, the fusion modules are the active ingredient rather than the specific pyramid configuration.","A fairer comparison would re-run all baselines inside FF-PNet's own preprocessing and training pipeline; the reported margins, such as 2.3 points over DualPRNet++, may shrink or shift if the original papers used different label definitions or data splits.","The paper's claim that convolutional encoders suffice suggests the transformer advantage in registration may be concentrated in the decoder. An exchange experiment with a transformer encoder and a CNN decoder, and vice versa, would isolate where the capacity actually matters."],"forward_implications":["Transformer or attention blocks are not required for high-accuracy brain registration; a CNN encoder plus the two fusion modules is enough to set the best reported Dice among the compared methods on LPBA and OASIS.","The modules are described as portable additions to the decoder, so any pyramid or U-Net registration framework adopting RFFM and RDFFM could expect a similar accuracy boost without adding attention machinery.","The layer ablation implies a three-level pyramid with fusion captures mid-level spatial correspondences better than a deeper four-level design, so adding depth is not the obvious path to further gains when these modules are active.","The full model is reported at roughly 8.9 million parameters, about a quarter of the four-layer variant with nearly identical Dice, which makes the design more practical for deployment."],"supporting_citations":[{"why":"VoxelMorph, the primary CNN-based U-Net registration baseline that FF-PNet is compared against on both datasets.","marker":"[12]"},{"why":"TransMorph, a hybrid CNN-transformer registration network whose reported Dice and HD95 numbers serve as transformer-baseline comparisons.","marker":"[14]"},{"why":"DualPRNet++, a pyramid registration baseline whose DSC margin of 2.3 points over FF-PNet's results motivates the multi-scale fusion design.","marker":"[16]"},{"why":"Im2Grid, a pyramid baseline with neighborhood attention, used as a comparison in the LPBA table.","marker":"[17]"},{"why":"PIVIT, a pyramid-iterative vision transformer baseline included in both dataset comparisons.","marker":"[19]"},{"why":"RDN, the recursive decomposition network whose LPBA DSC of 0.720 is the closest pyramid baseline to FF-PNet's 0.726.","marker":"[27]"},{"why":"H-ViT, the hierarchical vision transformer baseline that FF-PNet outperforms on all three OASIS metrics.","marker":"[28]"}],"fun_headline_variants":["Plain CNN pyramid with dual fusion beats transformer registration","Dual fusion modules let a plain CNN rival transformer registration","Pyramid with residual fusion beats attention models on brain MRI","CNN-only encoder, fusion decoder: new SOTA for brain registration","Dual fusion pyramid nets outdo transformers without attention"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy gains come from comparing FF-PNet's own runs with numbers copied from the original papers, assuming the same preprocessing, training setup, and label definitions; if those differ, the improvement could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Plain CNN pyramid with dual fusion beats transformer registration","Dual fusion modules let a plain CNN rival transformer registration","Pyramid with residual fusion beats attention models on brain MRI","CNN-only encoder, fusion decoder: new SOTA for brain registration","Dual fusion pyramid nets outdo transformers without attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3232,"prompt_tokens":869,"completion_tokens":2363,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":2283}},"tokens_in":485,"tokens_out":2363,"duration_ms":17530,"temperature":1.0,"reasoning_tokens":2283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:16:47.475799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VoxelMorph, TransMorph, and DualPRNet++ through FF-PNet's own training code, data splits, and hyperparameters on LPBA40 and OASIS, then compare Dice. If the re-run baselines match or exceed FF-PNet's 0.726 and 0.884 within one standard deviation, the claim of consistent improvement is not supported.","supporting_citations":[{"cited_title":"Dual- stream pyramid registration network,","cited_arxiv_id":null,"evidence_quote":"DualPRNet++, a pyramid registration baseline whose DSC margin of 2.3 points over FF-PNet's results motivates the multi-scale fusion design."},{"cited_title":"Coor- dinate translator for learning deformable medical image registration,","cited_arxiv_id":null,"evidence_quote":"Im2Grid, a pyramid baseline with neighborhood attention, used as a comparison in the LPBA table."},{"cited_title":"Pivit: Large deformation image registration with pyramid-iterative vision transformer,","cited_arxiv_id":null,"evidence_quote":"PIVIT, a pyramid-iterative vision transformer baseline included in both dataset comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RDN, the recursive decomposition network whose LPBA DSC of 0.720 is the closest pyramid baseline to FF-PNet's 0.726."},{"cited_title":"Recursive decomposition network for deformable image registration,","cited_arxiv_id":null,"evidence_quote":"H-ViT, the hierarchical vision transformer baseline that FF-PNet outperforms on all three OASIS metrics."}],"review_version":1}