{"id":"ccd70e50-2534-4be1-8eb3-614ee9429269","arxiv_id":"2501.10131","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A self-supervised pretraining method that aligns global and local patch embeddings via composition and decomposition improves transfer to medical imaging tasks.","lead":"ACE is a self-supervised training method for medical images that learns both large-scale and small-scale anatomical features by composing and decomposing image patches. It reports better few-shot and fine-tuned performance than several existing self-supervised baselines across chest X-ray and fundus tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Local consistency relies on exact patch indices after resizing and augmentation; the paper never states the augmentations used on the local branch or how the target matrix is aligned when resized crops change the grid, so the central anatomical-correspondence premise is unverifiable as written.","rationale":"In good faith, the paper proposes a coherent and well-motivated SSL objective. The strongest claim, that grid-wise cropping plus composition/decomposition yields anatomically consistent embeddings with superior transfer, is directly tied to Eq. 3-4 and the pseudo-code in Algorithm 1. The load-bearing premise is that the Gaussian target matrix can be computed from grid indices alone after augmentation. The manuscript contains two concrete contradictions: Sec. 3.1 says a 1024x1024 input with 32x32 patches (a 32x32 grid), while Sec. 4 and Appendix A.3 train on 448x448 images with 14x14 patch embeddings; and the augmentation transforms T1/T2 are never enumerated for the local branch, despite the pseudo-code applying augment(C1) and augment(C2) before computing losses. If flips or random-resized crops are in those transforms, the index-based matrix in ComputeLoss is misaligned for a substantial fraction of patches, so the model is trained on wrongly paired tissue. This is a correctness risk, not a novelty dispute: the proposed mechanism could still work, but as written it is not verifiable or reproducible. The missing Adam-v2 benchmark and qualitative property analyses are secondary; they affect the strength of the empirical comparison but not the validity of the mechanism itself. The reader's weakest_assumption identifies the same issue, so agreement is 'agree'. Reasonable verdict: conditional acceptance pending release of code and an explicit specification of the augmentations used on the local branch, with a rerun or analysis verifying that targets remain aligned. A hard reject would be too strong because the concern is testable and the method may be correct once the ambiguity is resolved; unconditional acceptance is too strong because the central mechanism is currently unverifiable.","tokens_in":20378,"tokens_out":2561,"duration_ms":22150,"concrete_test":"Obtain the released pretraining code (the paper promises GitHub.com/JLiangLab/ACE) and implement a minimal 'alignment audit' for the local branch: log the actual x,y patch coordinates of O1 and O2 after the two random crops, resizing, and every applied augmentation for a few thousand sampled crops, and compute the fraction of 2x2-to-1 patch pairs for which the Gaussian target Eq. 4 is aligned to the true anatomical correspondence (e.g., by comparing against the known grid offsets before augmentation). If this fraction is materially below 1 for any augmentation in the pipeline, rerun pretraining: (i) with the same augmentations but targets built from the true post-augmentation coordinates, and (ii) with flips/random-resized-crops removed from the local branch only.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that ACE produces anatomically consistent local embeddings via precise grid-based composition/decomposition (Sec. 3.3). The mechanism requires that, after resizing and augmentation, a 2x2 block of student tokens corresponds to exactly one teacher token at a known grid offset, so the Gaussian target in Eq. 4 is correct. The paper does not specify whether the local branch's augmentations (T1, T2 in Sec. 3.2 and augment(C1), augment(C2) in Algorithm 1) include flips, random-resized crops, or other transforms that break index-based correspondence; in fact, the method description says 'The resized crops C1 and C2 are added different augmentations x = T1(C1), x' = T2(C2)' without disambiguating global vs. local branches. Compounding this, Sec. 3.1 states the input image is 1024x1024 with 32x32 patches (a 32x32 grid), while Sec. 4 and Appendix A.3 describe training at 448x448 patch embeddings with 14x14 grids; Appendix A.3 says the composer concatenates 'adjacent embeddings' without defining adjacency after augmentation. Algorithm 1's ComputeLoss builds T from indices idx1 = nonzero(O1), idx2 = nonzero(O2) with no coordinate alignment step, so if either O1/O2 coordinates or the patch grids are not consistently aligned after resizing, the target is wrong. The reader flagged this and it is indeed load-bearing: if the target matrix assigns correct labels to wrong patches, the model is trained to align mismatched anatomy, and the claimed emergent cross-patient correspondence (Sec. 5.2) could be spurious or merely inherited from global statistics. A second-order concern is that baselines with the same backbone and pretraining data are not all reported for the exact same protocol, and Adam-v2, the closest compositional prior work, is only cited, not benchmarked; however, the augmentation/grid ambiguity is the deepest issue because it directly affects what the loss optimizes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ACE, a self-supervised pretraining method for medical images (primarily chest X-rays) that combines a global consistency branch, which aligns averaged patch embeddings of overlapping grid-based crops in a DINO-style student--teacher setup, with a local consistency branch that enforces composition and decomposition relationships between patch embeddings through Gaussian-smoothed correspondence matrices. The authors claim that ACE learns anatomically consistent embeddings, exhibits emergent properties such as cross-patient anatomical correspondence and symmetry, and transfers better than existing SSL baselines to classification, segmentation, few-shot, and keypoint-detection tasks. The method is evaluated on six downstream datasets with ViT-B and Swin-B backbones, ablations show that each loss component contributes, and code and pretrained models are promised for release.","tokens_in":20751,"tokens_out":7892,"duration_ms":70450,"significance":"If the claims hold, ACE is a valuable contribution to medical image SSL: it explicitly models hierarchical part--whole structure, a prior that most SSL methods ignore, through an intuitive grid-wise cropping scheme and two complementary losses. The paper has several strengths: a clear and coherent two-branch objective; a broad downstream evaluation spanning classification, segmentation, few-shot transfer, and landmark detection on two backbones; an ablation study in Fig. 9 showing that each loss term improves performance; statistical testing (two-sample t-test) reported in Table 1; and a commitment to releasing code and pretrained models. The main uncertainties are whether the local correspondence mechanism is correctly implemented as described and whether the claimed superiority over the closest prior method, Adam-v2, actually holds, since that baseline is not evaluated anywhere in the paper.","major_comments":[{"comment":"The local consistency branch hinges on precise 4-to-1 patch matching between C1 and C2 after resizing and augmentation, but the paper never specifies the augmentations applied to the local branch (T1 and T2 in Sec. 3.2; augment(C1) and augment(C2) in Algorithm 1). If these augmentations include flips or random-resized crops, the index-based target in Eq. (4) is wrong. Additionally, Sec. 3.1 describes 1024x1024 input with 32x32 patches (a 32x32 grid), while Sec. 4 and Appendix A.3 describe pretraining at 448x448 with a 14x14 grid, and Algorithm 1's ComputeLoss builds the target from idx1=nonzero(O1) and idx2=nonzero(O2) without any coordinate-alignment step. Please specify the augmentation pipeline for the local branch, describe how correspondences are recomputed when the two crops have different grid resolutions, and reconcile the input-size discrepancy. Without this, the central anatomical-correspondence premise is unverifiable and the local loss may be training the model to align mismatched tissues.","section":"Section 3.3 / Eq. (4) / Algorithm 1"},{"comment":"The closest prior work on compositional and decompositional anatomy SSL, Adam-v2 [46], is cited in the Related Work section as 'the latest research... exploiting learning from composition and decomposition in a hierarchical way,' but Adam-v2 is not included in any of the experiments: it is absent from Table 1, the few-shot experiments in Fig. 8, the ablation in Fig. 9, and the property analyses in Secs. 5.1 and 5.2. Since ACE's contribution is specifically the combination of global consistency with local composition/decomposition, the lack of a head-to-head comparison with Adam-v2 leaves the paper's central competitiveness claim unsubstantiated against the most relevant baseline. Please add Adam-v2 to the comparison tables and property analyses, or provide a specific technical justification for its exclusion.","section":"Related Work / Table 1"},{"comment":"The 'emergent' properties are supported by t-SNE visualizations in Fig. 6 and a single mean landmark error of 61 pixels in Fig. 7, with no error bars, no statistical tests, and no comparison to any baseline feature extractor under the same zero-shot protocol. As written, these results do not demonstrate that ACE's embeddings are uniquely anatomically organized; for example, the 61-pixel mean error could plausibly be matched by other SSL features on this task, especially given the high degree of consistency in chest X-ray anatomy. Please report quantitative accuracy (e.g., success rate within a pixel tolerance), the error distribution across the 13 landmarks, and direct comparisons to baselines on the same unsupervised landmark-matching task, preferably with multiple seeds.","section":"Section 5.2 / Figs. 6 and 7"},{"comment":"The Pytorch pseudo-code contains implementation errors that make it irreproducible as written: loss_decomp is computed with s1 and t2 instead of the decomposition branch's s2 and t1; the matching matrix is computed with torch.mul (element-wise multiplication) rather than the matrix multiplication required by Eq. (3); and the comments contain typos (e.g., 'input composer head' for the decomposer, 'flattern'). Please correct the pseudo-code so that it matches the main-text equations, and proofread the supplementary material for such errors.","section":"Appendix A.1 / Algorithm 1"}],"minor_comments":[{"comment":"The notation '4482' and '10242' appears to be a formatting artifact for 448x448 and 1024x1024; please fix these superscripts for clarity (e.g., 448^2, 1024^2).","section":"Sections 4 and 5.2"},{"comment":"The caption describes t-tests and blue highlighting, but the typeset table does not show any highlighting and no p-values or effect sizes are reported; please include the full statistical results either in the table or in the supplementary material so the 'no statistically significant difference' claims can be verified.","section":"Table 1 caption"},{"comment":"The KDE distributions for compositionality are shown only qualitatively; please report the mean and standard deviation of the cosine similarities for each method so the 'narrower and taller' claim can be assessed numerically.","section":"Figure 3"},{"comment":"The retrieval experiment reports ACE's accuracy (94.37%) in the text but no baseline numbers; please provide a full comparison table or include the baseline numbers directly in the figure.","section":"Figure 5"},{"comment":"Please clarify how the two random crops C1 and C2 are sampled from the 32x32 grid (e.g., random offsets, fixed relative positions) and how the overlap masks O1 and O2 are defined; this will help readers understand the correspondence construction that the method relies on.","section":"Section 3.1"},{"comment":"The operator '⊕' used for average pooling is not formally defined; please provide a precise definition to avoid confusion with the later uses of matrix notation.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the downstream transfer results are plausible, but the paper needs a substantive revision. The most serious issue is the unspecified local-branch augmentation and the unresolved grid-size discrepancy (1024 vs. 448), which makes the central patch-correspondence mechanism unverifiable as written; this is fixable by specifying the augmentations and adding an explicit coordinate-alignment step. The missing Adam-v2 comparison is also a major empirical gap, as Adam-v2 is the closest published method in exactly the dimension ACE claims as its novelty. The pseudo-code errors in the appendix are concerning for reproducibility but are easy to correct. If the authors resolve the correspondence issue, add the Adam-v2 baseline, and strengthen the emergent-property analyses with quantitative comparisons, the paper could become a strong contribution to medical image SSL."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ACE is a solid, well-motivated SSL paper for medical images, and the main thing to know is that the central idea—pairing DINO-style global consistency with local composition/decomposition losses on grid-aligned crops—is genuinely new relative to PEAC and Adam-v2. The fine-tuning results are not dramatic but are consistent across two backbones and several datasets, and the ablation in Fig. 9 shows each loss term helps. Credit is also due for the broad evaluation: six datasets, classification and segmentation, few-shot settings, and a serious attempt to test learned and emergent properties.\n\nThe soft spots are real but mostly fixable. Adam-v2, the closest prior work on composability and decomposability, is cited but never benchmarked; that is a conspicuous omission. Several property analyses (compositionality, decompositionality, retrieval, t-SNE landmark separation) are single-run numbers or visualizations without error bars or quantified baselines in every case. Most important for reproducibility: the paper does not specify the augmentation transforms used on the local branch, and the text moves between 1024x1024/32x32 and 448x448/14x14 grids. The stress-test worry—that flips or random-resized crops would break the index-based correspondence matrix in Eq. 4—is a fair reading of the text as written. If the local branch uses only non-spatial augmentations, the grid-wise crop design keeps the patches aligned and the concern disappears. But the authors need to state this explicitly. As written, the central anatomical-correspondence premise is not verifiable, which is grounds for requesting a revision, not a rejection.\n\nThis paper deserves a serious referee. It is aimed at the medical-imaging SSL community and at people working on part-whole representation learning. I would want the authors to add the Adam-v2 comparison, report error bars for the property analyses, and clarify the augmentation and grid pipeline. The promised code also needs to be actually available. If those points come through, the contribution is a useful one.","headline":"A solid, well-motivated medical SSL method with a genuinely new grid-based composition/decomposition loss; the evaluation is broad, but the missing Adam-v2 baseline and unspecified local-branch augmentations need fixing before the central correspondence claim is verifiable.","tokens_in":21405,"tokens_out":4106,"would_cite":true,"duration_ms":39845,"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":"ACE pretrains medical-image encoders to respect part-whole anatomy, and the resulting embeddings transfer to downstream classification, segmentation, and few-shot tasks.","keywords":["self-supervised learning","medical image analysis","compositionality","decompositionality","anatomically consistent embeddings","chest X-ray","few-shot learning","student-teacher model"],"falsifier":"Pretrain ACE with flips or random-resized crops that break grid alignment in the local branch while keeping everything else fixed; if decompositionality accuracy (the f(X) - f(X_excised) ≈ f(C) test) drops sharply or downstream transfer degrades, the precise-patch-matching premise is load-bearing. Alternatively, inspect the learned matching matrix M_comp: if high correlation values appear far from the Gaussian-smoothed diagonal, the model is not actually honoring spatial alignment.","tokens_in":20148,"feed_emoji":"🩻","tokens_out":5960,"duration_ms":52872,"temperature":0.7,"pith_summary":"The paper argues that medical images carry a composable and decomposable anatomical structure that existing self-supervised learning methods ignore. It introduces ACE, a self-supervised pretraining method that learns global consistency between overlapping crops and local consistency through composition and decomposition losses. If the claim holds, ACE gives medical imaging models a representation that transfers better to classification, segmentation, few-shot, and landmark tasks, which matters when labeled data is scarce. The paper reports this across six datasets and two backbones, with an ablation showing each loss component contributes.","feed_headline":"ACE learns anatomy by composing and decomposing image patches","feed_subtitle":"Pretrained on unlabeled chest X-rays, it beats SSL baselines on classification, segmentation, and few-shot tasks.","key_machinery":"The load-bearing mechanism is grid-wise image cropping: an input image is divided into a 32x32 patch grid, and two random crops of 14x14 and 28x28 patches are extracted so that every large-crop patch overlaps four small-crop patches in exact spatial correspondence. A composer head merges 2x2 student patch embeddings into one, a decomposer head splits one patch embedding into four, and CLIP-style cross-correlation matrices with Gaussian-smoothed targets supervise the matching. A student-teacher EMA setup ties the global and local branches together. This replaces approximate nearest-neighbor patch matching with precise, anatomy-aligned matching.","core_discovery":"ACE's central claim is that training a student-teacher network to respect part-whole relationships in standardized medical images produces anatomically consistent embeddings, and that this consistency transfers to downstream tasks. Concretely, the paper trains the embedding of a whole patch to be consistent with the aggregated embeddings of its sub-patches (composition) and trains sub-patch embeddings to be consistent with the decomposition of the whole-patch embedding (decomposition), using precisely matched grid-aligned crops. The paper reports that ACE outperforms a diverse set of SSL baselines on ChestX-ray14, Shenzhen, RSNA, JSRT, ChestX-Det, and SIIM, and that it shows emergent cross-patient landmark correspondence and useful few-shot performance, including over 91% of full-data heart-segmentation Dice from two labeled samples.","pith_inferences":["The paper tests ACE only on 2D projection images; if the spatial-alignment premise holds, the same grid-wise composition/decomposition losses could extend to 3D volumes such as CT and MRI, where sub-volume part-whole structure is even more explicit.","The Gaussian-smoothed target matrix is a soft spatial prior; replacing it with a hard one-hot target or a learned soft assignment would isolate how much of ACE's gain comes from the spatial prior versus the composition/decomposition objective itself.","The measured decompositionality property, where removing a region from an image yields an embedding difference close to the embedding of the excised region, suggests ACE features could support localized image editing or disease attribution, a use the paper only hints at through Grad-CAM visualizations.","The paper's local branch assumes augmentation preserves grid alignment; a testable consequence is that strong flips or severe random-resized cropping during pretraining should degrade the composition/decomposition accuracy and downstream transfer."],"forward_implications":["Pretrained ACE weights can be fine-tuned for thoracic disease classification and segmentation with substantially less labeled data; the paper reports 2-shot heart segmentation reaching over 91% of full-data performance.","The embedding space supports unsupervised cross-patient anatomy correspondence, meaning the same model can retrieve or match anatomical landmarks across different patients without fine-tuning.","Because the method only assumes standardized imaging protocols, it should transfer to other modalities with consistent anatomy; the paper demonstrates this on fundus photography.","The composition and decomposition losses provide a training signal that bridges large pathologies and small tissue anomalies, which should improve weakly supervised localization of small findings such as nodules."],"supporting_citations":[{"why":"Supplies the student-teacher EMA framework and global-consistency objective that ACE builds on.","marker":"[12]"},{"why":"Hinton's part-whole hierarchy paper motivates the composition and decomposition losses.","marker":"[22]"},{"why":"Prior work on hierarchical composition/decomposition in anatomy, which ACE extends from global patterns to local patches.","marker":"[46]"},{"why":"Prior local anatomical consistency method with precise matching that considered only positive pairs, which ACE extends with unpaired-patch separation.","marker":"[66]"},{"why":"Global-feature SSL baseline used in comparisons and property analyses.","marker":"[6]"},{"why":"Provides the patch-decomposition evaluation setup used to measure feature compositionality.","marker":"[24]"},{"why":"Supplies the cross-correlation matrix formulation used for local consistency matching.","marker":"[42]"}],"fun_headline_variants":["ACE: SSL that composes and decomposes medical images","Anatomy-aware SSL via patch composition and decomposition","Compose-decompose pretraining beats SSL baselines on X-rays","ACE learns anatomy by respecting part-whole structure","Part-whole consistency: a new SSL recipe for medical imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The local consistency branch assumes that, after resizing and random augmentation, a grid position in one crop still corresponds to the same anatomical grid position in the other crop, so the Gaussian-smoothed target matrix pairs the right patches.","fun_headline_variants_meta":{"raw":{"variants":["ACE: SSL that composes and decomposes medical images","Anatomy-aware SSL via patch composition and decomposition","Compose-decompose pretraining beats SSL baselines on X-rays","ACE learns anatomy by respecting part-whole structure","Part-whole consistency: a new SSL recipe for medical imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1514,"prompt_tokens":890,"completion_tokens":624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":543}},"tokens_in":506,"tokens_out":624,"duration_ms":6784,"temperature":1.0,"reasoning_tokens":543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:24:03.622710+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pretrain ACE with flips or random-resized crops that break grid alignment in the local branch while keeping everything else fixed; if decompositionality accuracy (the f(X) - f(X_excised) ≈ f(C) test) drops sharply or downstream transfer degrades, the precise-patch-matching premise is load-bearing. Alternatively, inspect the learned matching matrix M_comp: if high correlation values appear far from the Gaussian-smoothed diagonal, the model is not actually honoring spatial alignment.","supporting_citations":[{"cited_title":"Bootstrap your own latent-a new approach to self-supervised learning","cited_arxiv_id":null,"evidence_quote":"Supplies the student-teacher EMA framework and global-consistency objective that ACE builds on."},{"cited_title":"How to represent part-whole hierarchies in a neural network","cited_arxiv_id":null,"evidence_quote":"Hinton's part-whole hierarchy paper motivates the composition and decomposition losses."},{"cited_title":"Representing part-whole hierarchies in foundation models by learning localizability composability and decomposability from anatomy via self supervision","cited_arxiv_id":null,"evidence_quote":"Prior work on hierarchical composition/decomposition in anatomy, which ACE extends from global patterns to local patches."},{"cited_title":"Learning anatom- ically consistent embedding for chest radiography","cited_arxiv_id":null,"evidence_quote":"Prior local anatomical consistency method with precise matching that considered only positive pairs, which ACE extends with unpaired-patch separation."},{"cited_title":"Emerg- ing properties in self-supervised vision transformers","cited_arxiv_id":null,"evidence_quote":"Global-feature SSL baseline used in comparisons and property analyses."},{"cited_title":"Towards foundation models learned from anatomy in medical imaging via self-supervision","cited_arxiv_id":null,"evidence_quote":"Provides the patch-decomposition evaluation setup used to measure feature compositionality."}],"review_version":1}