{"id":"465a018f-c77c-4b97-825c-c1e28ced42ad","arxiv_id":"2502.05091","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A decomposed 3D convolution encoder for chest CT reaches competitive pathology detection and image-text retrieval with far fewer parameters and FLOPs than transformer or full 3D convolution baselines.","lead":"DCFormer is a 3D image encoder that replaces each full 3D convolution with three simpler 1D convolutions along the depth, height, and width axes, cutting compute significantly for chest CT scans. Trained in a CLIP-style vision-language setup on 50,188 paired CT volumes and radiology reports, it reaches competitive pathology detection and image-text retrieval with far fewer parameters than transformer or 3D-convolution baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 7's additive sum of three axis-wise 1D convolutions is a rank-limited spatial operator; without an ablation against a full 3D depthwise convolution, the claim that decomposition preserves spatial information is unverified.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing premise: Eq. 7's additive combination of three axis-wise 1D convolutions is assumed to capture the cross-axis interactions that a full 3D convolution would provide, with no theoretical guarantee and no ablation against a full 3D depthwise convolution. My stress-test sharpens this into a rank argument: the summed operator has at most 3k degrees of freedom per channel versus k^3 for a full kernel, so the spatial receptive field is a cross shape rather than a full cube. This is a genuine capacity restriction, and whether it matters for CT-RATE is an empirical question the paper does not directly answer. The absence of error bars and the small, sometimes non-monotonic differences in Tables 1-3 further weaken the empirical support for the 'consistently outperforms' claim. I do not see an internal inconsistency that would warrant rejection; the efficiency math is correct and the architecture is clearly described. The appropriate disposition remains CONDITIONAL: the central claim needs the decomposition ablation and variance reporting before the representation-preservation claim can be accepted. Since this is the same condition the reader already attached, the verdict is unchanged.","tokens_in":12030,"tokens_out":5680,"duration_ms":61004,"concrete_test":"Train the DCFormer-naive CLIP pipeline on CT-RATE with the DecompConv3D token mixer replaced by a standard 3D depthwise convolution, matching total model parameters and FLOPs (e.g., full depthwise k=3 with appropriately adjusted channel widths, or full depthwise k=7 with reduced channels). Run at least 3 seeds and report zero-shot F1 and R@50. If full 3D depthwise matches or beats the decomposed version at the same budget, the central claim that factorization preserves spatial information fails. If the decomposed version retains its advantage, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decomposition in Eq. 7 is not merely an efficiency trick; it defines the spatial operator. The effective kernel seen by a token is the sum of three axis-aligned 1D-depthwise kernels, so at each layer the token mixer has support only on the H, W, and D axes and zero on all off-axis positions. A full 3D depthwise kernel of size k has k^3 degrees of freedom per channel; the decomposed version has 3k, i.e., rank at most 3, so it cannot represent cross-axis correlations within a single layer. Whether stacked layers plus nonlinearities compensate for this is an empirical question, and the paper provides no ablation against a standard 3D depthwise convolution at a matched parameter/FLOP budget. The comparisons in Tables 1-3 use strong baselines, but they do not include the natural control for the central architectural claim. The reported margins are also small and no error bars or seed variance are given; some baseline numbers exceed DCFormer on specific metrics (e.g., TransUNet-naive has higher R@10 and R@50 than DCFormer-naive in Table 3, and ConvNeXt-tiny has higher zero-shot accuracy than DCFormer-tiny in Table 1), so the abstract's 'consistently outperforms' is overstated. Thus the efficiency gain is certain, but the representation-preservation claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DCFormer is a 3D vision encoder that replaces the token mixer in a MetaFormer block with three parallel axis-wise 1D depthwise convolutions along height, width, and depth, whose outputs are normalized and summed (Eq. 7). This decomposition reduces parameters from Ck^3 to 3Ck and FLOPs from 2CHWDk^3 to 6CHWDk. The encoder is hierarchical with four stages, and it is integrated into a CLIP framework with a CXR-BERT text encoder. The paper evaluates this model on the CT-RATE dataset for zero-shot and fine-tuned detection of 18 pathologies, and for image-text retrieval, comparing against CT-ViT, ViT, ConvNeXt, PoolFormer, and TransUNet. The central claim is that the decomposed convolution preserves spatial information while greatly reducing cost, and that DCFormer consistently outperforms the compared encoders.","tokens_in":12376,"tokens_out":7651,"duration_ms":77277,"significance":"The efficiency analysis is correct and clearly presented: the parameter and FLOP formulas for the decomposed convolution are properly derived, and the hierarchical design is well specified. The paper also ships code and uses a unified CLIP training framework across all compared encoders, which is a strength for fair comparison. If the central architectural claim were established, DCFormer would be a useful and practical 3D vision encoder for medical vision-language models. However, the current evidence does not yet establish that the decomposed convolution preserves the spatial information captured by a full 3D convolution, and the reported performance advantages are often small and unsupported by repeated runs or significance testing. The significance is therefore conditional on additional controlled experiments.","major_comments":[{"comment":"The central architectural assumption — that X' = X + Norm_h(X*h) + Norm_w(X*w) + Norm_d(X*d) is an adequate replacement for a 3D token mixer — is not tested against the natural control of a full 3D depthwise convolution at a matched parameter/FLOP budget. The effective per-layer operator has support only on the H, W, and D axes, so it cannot represent cross-axis correlations within a single layer. Whether stacking layers and nonlinearities recovers this capacity is an empirical question, and the paper provides no ablation against a standard 3D depthwise convolution (or a full 3D convolution) to support the claim that spatial information is preserved. This is the load-bearing experiment for the core design choice and is currently missing.","section":"Formulation of the DCFormer block, Eq. (7)"},{"comment":"All reported results are single runs with no error bars, no multiple seeds, and no statistical significance testing. The margins over strong baselines are often small or reversed. For example, in Table 1 the zero-shot accuracy of DCFormer-tiny (62.0%) is lower than CT-ViT (62.9%) and ConvNeXt-tiny (62.5%), and DCFormer-nano (60.4%) is lower than ConvNeXt-nano (62.2%). In Table 3, TransUNet-naive achieves higher retrieval R@10 (2.36% vs. 2.20%) and R@50 (10.00% vs. 9.54%) than DCFormer-naive. Therefore the claim in the abstract and Discussion that DCFormer 'consistently outperforms' the compared state-of-the-art encoders is not supported by the reported data.","section":"Tables 1 and 3; Abstract and Discussion"},{"comment":"The retrieval protocol is under-specified. It is not stated whether R@k is computed against the full validation set, how many candidates are in the retrieval pool, or whether the reported numbers correspond to the validation split used in the other tasks. Because the retrieval task is one of the three main evaluation pillars, the experimental setup must be described precisely so that the results can be independently reproduced and interpreted.","section":"Image-Text Retrieval (Table 3)"}],"minor_comments":[{"comment":"The relationship between Eq. (7) and the MetaFormer recurrence in Eq. (2) is unclear: the pseudocode applies BatchNorm3d inside each of the three parallel branches but does not show a normalization applied to X before the three convolutions, while Eq. (7) explicitly normalizes the three convolution outputs. Please clarify whether the block applies a shared normalization to the input of the token mixer and how the per-branch norms relate to Norm1 in Eq. (2).","section":"Eq. (7) and Algorithm 1"},{"comment":"The claim that DCFormer has been 'independently validated' in Med3DVLM (ref. 33) is misleading because the author list of Med3DVLM overlaps with the present authors. Please describe this as additional validation by the same group, not an independent evaluation.","section":"Discussion, Med3DVLM validation"},{"comment":"The text states that kernel sizes are chosen from {13, 11, 9, 7}, but the architecture table and all reported experiments use kernel size 7 only. Please clarify whether larger kernels were tested or whether the set notation refers to a design choice that was not exercised in the experiments.","section":"Methods, kernel size notation"},{"comment":"Please state explicitly whether all baseline encoders (ViT, TransUNet, ConvNeXt, PoolFormer, CT-ViT) were trained with exactly the same training recipe, including the CLIP loss, batch size, learning rate, number of epochs, and absence of scheduling, and whether any per-model hyperparameter tuning was performed.","section":"Implementation Details"},{"comment":"There is a duplicated word in the sentence 'We compute the cosine similarity between between the CT image embedding and each of the text prompt embeddings.'","section":"Zero-Shot Multi-Abnormality Detection"},{"comment":"The novelty claim that DCFormer is 'the first 3D vision encoder to use decomposed convolutions to improve vision-language modeling' should be scoped more carefully and should cite prior work on factorized or axial convolutions, which are closely related to the proposed decomposition.","section":"Introduction and Related Work"},{"comment":"Table 3 uses the spelling 'naive' while the rest of the text uses 'naïve'; please unify the spelling.","section":"Table 3"},{"comment":"The Discussion claims that DCFormer 'can run on standard hospital hardware without requiring specialized infrastructure' and is suited for real-time applications, but the paper does not report inference time or latency. Consider adding latency measurements to support this claim.","section":"Discussion, clinical deployability"}],"recommendation":"major_revision","confidential_remarks":"The efficiency analysis and the clarity of the architecture description are solid, and the code release is commendable. The main problem is that the core architectural claim — that the decomposed convolution preserves the spatial information of a full 3D convolution — lacks a direct ablation. The missing control (full 3D depthwise convolution at matched parameters/FLOPs) is exactly the experiment needed to justify the design. In addition, the absence of error bars and the presence of metric reversals in Tables 1 and 3 mean the headline 'consistently outperforms' is not currently supported. I would ask the authors to add the matched-budget ablation, report multiple seeds or confidence intervals, specify the retrieval protocol, and temper the claims accordingly. The 'independent validation' statement about Med3DVLM should also be corrected, as it appears to be by the same group."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DCFormer does exactly what the title says: factorize each 3D depthwise convolution into three parallel 1D depthwise convolutions along H, W, D, and use that as the token mixer in a CLIP vision encoder for chest CT. The efficiency math checks out — parameters drop from Ck^3 to 3Ck, FLOPs from 2CHWDk^3 to 6CHWDk — and the architecture is described precisely enough to reimplement. Code is released. That part is solid.\n\nWhat's new is the application: nobody has put this particular factorization inside a 3D medical vision-language model. The hierarchical design and the comparison against ViT, ConvNeXt, PoolFormer, TransUNet, and CT-ViT on CT-RATE are useful. The authors also cite their own Med3DVLM as external validation, which is a bit incestuous, but it's a separate paper.\n\nThe soft spots are real but not fatal. First, no citation to R(2+1)D or P3D, so the novelty framing oversells a known idea. Second, and more important, there is no ablation against a standard 3D depthwise convolution at a matched parameter/FLOP budget. Eq. 7 sums three axis-wise responses after normalization — this is a rank-limited operator, and whether stacked layers and nonlinearities make up for it is an open question that the paper never tests. Third, the 'consistently outperforms' claim doesn't survive a close read of the tables: ConvNeXt-tiny beats DCFormer-tiny on zero-shot accuracy, and TransUNet-naive beats DCFormer-naive on recall@10 and @50. Finally, all numbers are single-run, no error bars, and the retrieval scores are tiny across the board, which makes the ranking less informative than it looks.\n\nI don't think the paper is wrong. The efficiency gain is certain, and the performance story is plausible. But the central representational claim — that the decomposed mixer preserves enough 3D structure to match full 3D depthwise convolution — is not empirically established. That's a fixable gap: add the control, run multiple seeds, and soften the language. I'd send this to review; the architecture is useful and the dataset setup is solid. A referee who asks for that ablation would do the paper a favor.","headline":"The efficiency gain is real and the architecture is clearly described, but the paper never runs the one control experiment that would support its claim that decomposed convolutions preserve 3D spatial information.","tokens_in":12841,"tokens_out":2510,"would_cite":false,"duration_ms":24532,"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":"The paper claims that factorizing each 3D convolution into three parallel 1D depthwise convolutions cuts the per-layer cost from $Ck^3$ to $3Ck$ parameters and from $2CHWDk^3$ to $6CHWDk$ FLOPs while matching or beating larger encoders on…","keywords":["3D vision-language model","decomposed convolution","depthwise separable convolution","CT-RATE","zero-shot pathology detection","CLIP","medical image analysis","efficient deep learning"],"falsifier":"Run the same CLIP training with everything identical except that the token mixer is a full 3D depthwise convolution matched in FLOPs or parameters, and compare zero-shot F1 on the 18 CT-RATE abnormalities; a clear win for the full mixer, or a sharp drop when any one of the three axis branches is removed, would show the additive decomposition is losing information that the reported tables cannot resolve.","tokens_in":11860,"feed_emoji":"🩻","tokens_out":9264,"duration_ms":91405,"temperature":0.7,"pith_summary":"This paper tries to establish that a single architectural substitution—replacing each 3D convolution with three parallel 1D depthwise convolutions along depth, height, and width—is enough to make a 3D vision-language encoder for medical CT efficient without sacrificing accuracy. The authors build a CLIP-style model whose image encoder is a hierarchical stack of such decomposed-convolution blocks, and train it on 50,188 chest CT volumes paired with radiology reports. In zero-shot and fine-tuned detection of 18 abnormalities, and in image-text retrieval, their smallest encoders match or beat much heavier baselines while using far fewer parameters and FLOPs. If the claim holds, the same encoder design could become a cheap drop-in image tower for volumetric medical vision-language systems.","feed_headline":"Three 1D convolution sweeps replace each 3D kernel in a CT encoder","feed_subtitle":"Kernel cost goes from cubic to linear while a small encoder matches or beats heavier 3D models on 18 pathology tasks.","key_machinery":"The load-bearing mechanism is the Decomposed Convolution, defined in Eq. 7 as $$X' = X + \\mathrm{Norm}_h(X \\ast_h) + \\mathrm{Norm}_w(X \\ast_w) + \\mathrm{Norm}_d(X \\ast_d),$$ where $X \\ast_h = \\mathrm{DWConv}_{k\\times1\\times1}(X)$, $X \\ast_w = \\mathrm{DWConv}_{1\\times k\\times1}(X)$, and $X \\ast_d = \\mathrm{DWConv}_{1\\times1\\times k}(X)$ are depthwise convolutions along height, width, and depth, each followed by its own normalization. The block inherits the MetaFormer layout—normalization, token mixer, channel MLP, residual connection—with the decomposed convolution in place of self-attention. This is the component that turns $Ck^3$ parameters into $3Ck$ and $2CHWDk^3$ FLOPs into $6CHWDk$, which is what makes large kernels (here 7) affordable in deep 3D networks.","core_discovery":"The central claim is that a 3D depthwise convolution can be replaced, as a token mixer, by three parallel 1D depthwise convolutions—one along each spatial axis—and still serve as the spatial backbone of a CLIP-style vision-language model. The replacement changes the per-layer cost from $Ck^3$ parameters and $2CHWDk^3$ FLOPs to $3Ck$ parameters and $6CHWDk$ FLOPs, turning cubic growth in kernel size into linear growth. On CT-RATE, DCFormer variants report the highest F1 scores in both zero-shot and fine-tuned detection of 18 chest abnormalities among the compared encoders, while the naïve variant uses 5.85 million parameters against CT-ViT's 101.1 million. The authors attribute the result to large-kernel decomposed convolutions giving broad spatial context inside a hierarchical, multi-scale encoder.","pith_inferences":["The effective receptive field of one decomposed block is a 3D cross—each axis is processed independently and the results are summed—so stacked layers must build up oblique or curved structure across multiple blocks; whether that is a strength or weakness for specific pathologies is not tested.","The same factorization transfers directly to any 3D modality where local features are roughly axis-aligned, such as MRI, PET, or volumetric video; the paper only evaluates chest CT, so the transfer is an extrapolation.","One could push kernel sizes to 31 or 51, as in large-kernel 2D work, at a cost that remains linear, turning the design into a test bed for how much receptive field 3D medical vision-language models actually need."],"forward_implications":["The encoder family scales from 0.92M to 15.1M parameters and 34 to 168 GFLOPs, so the same CLIP training and inference pipeline runs on much lighter hardware than CT-ViT's 101.1M parameters and 160.5 GFLOPs.","Because cost scales linearly with kernel size, the design makes large-kernel 3D convolutions practical; larger kernels than 7 are a natural next knob to turn.","A single linear layer on top of the frozen encoder gives fine-tuned multi-label detection an F1 score of 48.6% for the tiny variant, the best reported in the comparison.","The hierarchical four-stage structure preserves multi-scale features that a fixed large-patch tokenizer such as CT-ViT loses, which the authors connect to medical imaging performance."],"supporting_citations":[{"why":"Defines the contrastive image-text alignment loss and training paradigm that the DCFormer CLIP framework is built on.","marker":"[16]"},{"why":"Supplies the CT-RATE dataset, the CT-ViT and CT-CLIP baseline, the preprocessing protocol, and the 18-abnormality labels.","marker":"[26]"},{"why":"Introduces the 7x7 depthwise-convolution token mixer and the ConvNeXt baseline family that DCFormer is compared against.","marker":"[31]"},{"why":"Provides the MetaFormer backbone and the PoolFormer baseline; DCFormer's block structure follows this paradigm.","marker":"[32]"},{"why":"Defines the ViT baseline whose quadratic self-attention cost the paper argues against for 3D volumes.","marker":"[7]"},{"why":"Introduces depthwise separable convolutions, the efficiency concept that the decomposed 1D depthwise convolution extends to 3D.","marker":"[29]"},{"why":"Provides the domain-specific CXR-BERT text encoder used in the CLIP framework.","marker":"[30]"}],"fun_headline_variants":["Cubic to linear: decomposing 3D convs accelerates CT vision-language","Three 1D convolutions replace 3D kernels, beating heavier CT encoders","DCFormer's 1D decomposition scales 3D medical VLMs without the cubic cost","Linear kernel cost with 1D sweeps: DCFormer improves CT vision-language"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the assumption that summing three axis-wise 1D sweeps preserves the cross-axis spatial interactions a full 3D convolution would capture; if those interactions matter for a pathology, the efficiency gain comes at a representational cost.","fun_headline_variants_meta":{"raw":{"variants":["Cubic to linear: decomposing 3D convs accelerates CT vision-language","Three 1D convolutions replace 3D kernels, beating heavier CT encoders","DCFormer's 1D decomposition scales 3D medical VLMs without the cubic cost","Linear kernel cost with 1D sweeps: DCFormer improves CT vision-language"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001196,"raw_usage":{"total_tokens":4959,"prompt_tokens":996,"completion_tokens":3963,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":3873}},"tokens_in":612,"tokens_out":3963,"duration_ms":30698,"temperature":1.0,"reasoning_tokens":3873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T20:17:08.777293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same CLIP training with everything identical except that the token mixer is a full 3D depthwise convolution matched in FLOPs or parameters, and compare zero-shot F1 on the 18 CT-RATE abnormalities; a clear win for the full mixer, or a sharp drop when any one of the three axis branches is removed, would show the additive decomposition is losing information that the reported tables cannot resolve.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the contrastive image-text alignment loss and training paradigm that the DCFormer CLIP framework is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CT-RATE dataset, the CT-ViT and CT-CLIP baseline, the preprocessing protocol, and the 18-abnormality labels."},{"cited_title":"Metaformer is actually what you need for vision","cited_arxiv_id":null,"evidence_quote":"Provides the MetaFormer backbone and the PoolFormer baseline; DCFormer's block structure follows this paradigm."},{"cited_title":"Xception: Deep learning with depthwise separable convolutions","cited_arxiv_id":null,"evidence_quote":"Introduces depthwise separable convolutions, the efficiency concept that the decomposed 1D depthwise convolution extends to 3D."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the domain-specific CXR-BERT text encoder used in the CLIP framework."}],"review_version":1}