{"id":"86b883f4-fd2e-4820-bc8c-0af4bb08d535","arxiv_id":"2411.17845","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CABLD combines TPS-based registration with consistency losses and 1x1 random convolution augmentation to detect AFIDs landmarks from a single template, achieving low MRE on T1w and T2w MRI.","lead":"A new self-supervised method detects 32 brain landmarks in MRI scans using only one labeled reference template plus unlabeled scans, and reports strong accuracy across four datasets. It is potentially useful because it could cut the cost of expert landmark annotation and work on scans with different MRI contrasts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's SOTA claim fails on the key T2w generalization set: Table 2 shows CABLD MRE 3.99 mm vs ANTs(MI) 3.91 mm, and SDR@3mm 27.19% vs 35.00%; the paper reports no significance test for this comparison.","rationale":"The reader's verdict is CONDITIONAL, with one listed reason being that 'the paper overstates the T2w result.' My read agrees with that specific point and elevates it because it bears directly on the strongest claim: the abstract asserts SOTA performance on all four datasets, and the paper's own Table 2 provides a counterexample for the T2w dataset on the primary metric. This is not a disagreement with the field's consensus or a stylistic complaint; it is a checkable factual discrepancy between the stated claim and the reported numbers. The self-referential consistency-loss concern raised by the reader is also important, but it concerns the mechanism rather than the empirical result, and the ablation study shows the loss has a large effect, so the empirical outcome is less directly threatened. The T2w issue is more load-bearing: if the hypothesis test shows CABLD is not better than ANTs(MI) on MRE or SDR@3mm, the central sentence of the abstract is false as written. The paper still has value: the method is data-efficient, uses a single annotated template, and performs well on T1w datasets, with public code. Therefore the appropriate disposition remains CONDITIONAL, not REJECT, and since the reader already reached CONDITIONAL, I do not move the verdict. My concrete test would settle whether the abstract needs a precise qualifier or a stronger correction.","tokens_in":19611,"tokens_out":12285,"duration_ms":121378,"concrete_test":"Run a paired permutation or bootstrap test on per-subject mean radial errors between CABLD and ANTs(MI) on the 30 HCP-T2w scans, using the released code and the same preprocessing; report the p-value and 95% CI for the MRE difference, and likewise for SDR@3mm. If CABLD is not significantly better (or is significantly worse) on MRE or SDR@3mm, the abstract must be revised to say 'comparable to' or 'better on coarse thresholds' rather than 'outperforms SOTA in MREs and SDRs.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CABLD 'outperforms the state-of-the-art methods in terms of MREs and SDRs' across four datasets, including T2w scans. Table 2, the only T2w evaluation, contradicts this on the two most commonly reported metrics: CABLD has a higher (worse) mean radial error than ANTs(MI) (3.99 +/- 2.25 mm vs 3.91 +/- 2.19 mm) and a lower SDR at 3 mm (27.19% vs 35.00%), and it is also below MultiGradICON on SDR@3mm (33.33%). CABLD leads only on SDR@6mm and SDR@9mm. The text acknowledges the 3 mm deficit as 'slightly lower' but does not acknowledge the MRE deficit at all. Section 4.3 states that paired t-tests were used, but no significance result is reported for the CABLD vs ANTs(MI) comparison on HCP-T2w. Because T2w is the sole evidence for 'contrast-agnostic generalization to unseen contrasts,' the headline claim is not supported as stated. If the MRE difference is statistically significant, the abstract's universal SOTA claim is actively contradicted; even if it is not significant, 'outperforms' is misleading when the method is numerically worse on the primary metric for the primary generalization dataset.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CABLD, a self-supervised framework for 3D brain landmark detection that uses a single annotated template and unlabeled training scans. The method combines an inter-subject landmark consistency loss with an image registration loss, applies random-convolution contrast augmentation, and uses a curriculum-scheduled mixture of the two losses. Experiments are reported on four datasets (HCP T1w, OASIS, SNSX, and HCP T2w) using the AFIDs landmark protocol, with claims of state-of-the-art accuracy.","tokens_in":20022,"tokens_out":10837,"duration_ms":88460,"significance":"If the claims hold, the approach would substantially reduce annotation cost for landmark detection and offer a way to generalize to unseen MRI contrasts. The public code release, the use of the well-defined AFIDs protocol, and the breadth of evaluation datasets are strengths. However, the headline claims of universal state-of-the-art performance are only partially supported by the reported evidence, and several methodological details require clarification or correction.","major_comments":[{"comment":"The universal SOTA claim is contradicted by the T2w results. On HCP-T2w, CABLD has a higher (worse) MRE than ANTs(MI) (3.99±2.25 mm vs 3.91±2.19 mm) and a lower SDR at 3 mm (27.19% vs 35.00%; MultiGradICON also reaches 33.33%). CABLD is best only at SDR@6mm and SDR@9mm. Section 4.4 acknowledges the SDR@3mm deficit as 'slightly lower' but does not acknowledge the MRE deficit. Although Section 4.3 states that paired t-tests were used, no p-values are reported for these comparisons. Since HCP-T2w is the only evidence for contrast-agnostic generalization, the abstract's claim that CABLD 'outperforms the state-of-the-art methods in terms of MREs and SDRs' should be revised to a dataset-specific, significance-tested statement.","section":"Section 4.4, Table 2, Abstract"},{"comment":"The fully supervised baseline appears to be trained on the labeled test data. Section 4.1 defines the testing data as 122 scans from four sources and states that for each scan, 32 AFIDs landmarks were manually labeled. Section 4.3 says the supervised CNN is trained 'using the labeled datasets described in Sec. 4.1'. If the labeled test scans were used for training, the comparison with CABLD is invalid because the baseline had access to the test labels, making the claim that CABLD 'outperform[s] the supervised 3D CNN with statistical significance' misleading. Please clarify the exact data split and, if the supervised CNN was trained on test labels, retrain it with a disjoint train/test partition.","section":"Section 4.3 and Section 4.1"},{"comment":"The consistency losses are self-referential and may be degenerate. In Eq. (1), β* is chosen to minimize Σ_j (T_β(f(x_i;θ))(j) - P(j))^2 + λI; Eq. (4) then evaluates exactly the first part of that minimized objective at β*, i.e., L_consistency2 is the squared residual of the regularized TPS fit against P. For λ=0, TPS interpolation makes this term exactly zero for any prediction; for small λ, it measures only the residual of a fit that was designed to minimize it. L_consistency1 (Eq. 3) inherits the same issue because both warps map their predictions near P. The paper does not report the actual distribution for sampling λ (a 'log-uniform distribution ranging between 0 and 10' is not well-defined), nor any analysis of the gradients. The mechanism claimed for 'inter-subject landmark consistency' is therefore not established. Please reformulate the consistency loss so it provides an independent constraint (e.g., by penalizing predicted-landmark distances to template-space locations without using a warp fitted to those same predictions) or provide a rigorous analysis of why the current formulation is non-degenerate in practice.","section":"Section 3.2, Eqs. (1)-(5)"},{"comment":"The random convolution augmentation uses 1×1×1 kernels on single-channel input, which reduces to a per-voxel pointwise transformation (affine scaling plus bias followed by LeakyReLU) rather than a spatial convolution that mixes neighboring structures. The manuscript characterizes this as '3D convolution-based contrast augmentation' capable of modeling 'complex, non-linear intensity relationships'; as written, the augmentation is closer to a soft histogram transform. Because the contrast-agnostic claim rests partly on this augmentation, please clarify the operational effect of 1×1 random convolutions and temper the novelty statement accordingly.","section":"Section 3.4 and Section 4.2"}],"minor_comments":[{"comment":"Equation (1) writes the regularizer as λI, while the supplementary material defines the bending energy as I_T; please use consistent notation.","section":"Eq. (1) and Supplementary"},{"comment":"The phrase 'randomly sampled from a log-uniform distribution ranging between 0 and 10' is not mathematically well-defined because a log-uniform distribution cannot include 0; please specify the exact sampling range and base.","section":"Section 4.2"},{"comment":"There are several typos, including 'anotmical' and 'paramter', which should be corrected.","section":"Section 3.2"},{"comment":"In the ablation table, the parenthetical values such as '(-49.99)' are not labeled; please clarify that they are absolute reductions in MRE (mm) relative to the base model.","section":"Table 3"},{"comment":"The y-axis of Figure 3 should specify that MRE is in millimeters and state that lower values are better, to avoid ambiguity.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The train-on-test issue for the fully supervised baseline is a serious validity concern that should be addressed before the paper can be considered further. The unsupported universal SOTA claim in the abstract will likely draw immediate criticism from reviewers and readers. The degeneracy of the consistency loss is a conceptual problem that may require substantial reformulation or analysis. I recommend major revision with careful attention to these three points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The gist: CABLD is a genuinely useful self-supervised landmark detection framework that needs only one annotated atlas, and its T1w results are strong. But the paper claims more than it shows: on the only unseen-contrast test (HCP-T2w), it loses to ANTs(MI) on MRE and SDR@3mm, and the fully supervised baseline is trained on test labels. Still a paper worth reviewing.\n\nWhat is actually new: the 3D random convolution (1x1x1) contrast augmentation. That is simple, sensible, and a first for this task, and the ablation shows it is what makes T2w transfer work. The single-template consistency scheme (TPS warp to the atlas plus cross-subject consistency and a curriculum ramp) is a reasonable package, and the authors show it beats KeyMorph, BrainMorph, and registration-based baselines on three T1w datasets, with p-values. Releasing code and including downstream PD/AD experiments is a plus.\n\nThe load-bearing problem is the abstract's SOTA claim. Table 2 is the only T2w evidence: CABLD's MRE is 3.99 vs ANTs 3.91, and SDR@3mm 27.19% vs 35.00%. The text calls it 'comparable' and 'slightly lower,' which is honest, but the abstract says 'outperforms ... in terms of MREs and SDRs' - that is contradicted by its own table. Since T2w is the sole test of contrast-agnostic generalization, the headline needs to be qualified, and the paired t-test reported for this comparison.\n\nThe supervised CNN baseline is trained on the same labeled test sets (Sec 4.3 uses labels from Sec 4.1), which makes it an unfair ceiling. The fact that it still loses on T1w is encouraging, but the comparison should be rerun or explicitly disclosed.\n\nThere is also a conceptual issue with the consistency loss. Equations 1-4 fit the TPS warp to make the predicted landmarks land on the template, then L_consistency2 measures the distance between those warped predictions and the template. That distance is minimal by construction (for small lambda), so the term is largely self-referential. The real learning signal must be coming through L_registration, because the warp derived from bad landmarks produces a bad image-to-template match. If so, the ablation showing L_consistency alone drops MRE from 53 to 3.7 is doing more work than the loss definition seems to support. The authors should analyze or explain this rather than leave it as an implicit bootstrap.\n\nThis is a practical, label-efficient method for clinical MRI landmark detection/QC, and the T1w results are credible. The flaws are in reporting and one under-theorized loss, not in the core engineering. I'd send it to peer review with the expectation of substantial revision; worth a serious referee.","headline":"A practical self-supervised landmark detector with a genuinely nice contrast-augmentation trick, but the abstract over-claims on T2w and the supervised baseline is compromised.","tokens_in":20496,"tokens_out":4050,"would_cite":true,"duration_ms":35164,"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":"A self-supervised framework detects 32 brain landmarks in unseen MRI contrasts using only a single annotated template.","keywords":["brain landmark detection","self-supervised learning","magnetic resonance imaging","contrast-agnostic","random convolution","landmark consistency","image registration","AFIDs protocol"],"falsifier":"A reader could retrain CABLD on the same unlabeled T1w scans but replace the two consistency losses with direct regression to the 32 template landmarks on a small labeled subset; if direct supervision produces lower mean radial error, the paper's single-template sufficiency claim is weakened. Alternatively, re-running with the curriculum schedule inverted, putting consistency first and registration second, would test whether the reported convergence and accuracy actually depend on the adaptive ordering.","tokens_in":19446,"feed_emoji":"🧠","tokens_out":4680,"duration_ms":44078,"temperature":0.7,"pith_summary":"CABLD aims to detect standardized anatomical landmarks in 3D brain MRI without large labeled datasets. It trains a 3D CNN to predict the 32 AFIDs landmark positions using a single annotated template (ICBM152) plus unlabeled scans, by fitting thin-plate splines from each predicted landmark set to the template and checking whether the resulting warp aligns brain volumes. Two consistency losses, one across sampled subjects and one between subject and template, pull the predicted landmarks into a fixed protocol, while a 3D random-convolution augmentation teaches the model to ignore contrast differences. The paper reports mean radial errors around 3 to 5 mm across four datasets, including T2w scans never seen in training, and says the method outperforms classical registration, keypoint-based learning methods, and a fully supervised CNN.","feed_headline":"One annotated template is enough to detect 32 brain landmarks","feed_subtitle":"Trained only on T1w scans, CABLD still detects the 32 AFIDs landmarks in T2w images with minimal error.","key_machinery":"The load-bearing mechanism is a differentiable thin-plate spline: at every training step the network's current predicted landmarks are used to fit a TPS warp to the template landmarks (with a regularization parameter drawn from a log-uniform distribution), and the same warp feeds both the registration loss and the landmark consistency losses. A second component is 3D random convolution with 1x1x1 kernels and LeakyReLU activations, which creates many artificial contrast variants from each T1w scan so that the model learns contrast-invariant landmark features while keeping the monomodal MSE loss usable. An adaptive mixing coefficient alpha, growing from near zero to one over training, shifts emphasis from registration to consistency in a curriculum-learning style.","core_discovery":"The paper's central claim is that registration and landmark consistency reinforce each other well enough to bootstrap a clinically meaningful landmark detector from a single annotated reference. The registration loss gives the network anatomical context by warping each subject scan to the template and comparing voxel intensities, while the two consistency losses force the predicted landmark positions to agree in template space across subjects and with the template itself. With an adaptive schedule that starts with registration and then shifts to consistency, the model learns to place the 32 AFIDs landmarks in T1w scans and generalizes to T2w scans without any T2w labels. The authors argue that unlike keypoint methods driven purely by registration similarity, CABLD produces landmarks that follow a predefined anatomical protocol rather than landmarks that merely happen to align scans.","pith_inferences":["Because the random-convolution augmentation is purely intensity-based, the same training recipe could plausibly transfer to other contrasts such as FLAIR or PD-weighted MRI without collecting paired multi-contrast data; this is a direct test a reader could run.","The implicit coordinate system created by the template-anchored consistency loss might make CABLD useful as a fast quality-control check for subject-to-template registration, since the landmarks are protocol-consistent and cheap to compute.","The curriculum schedule is itself an empirical claim: someone could vary the alpha schedule or remove the registration term after warmup and measure whether convergence and final MRE change, which would reveal how necessary the two-stage balance really is.","If pathological brains preserve the same gross anatomy, the method may also work on diseased scans without retraining, but that requires validation beyond the healthy subjects used for the main tables."],"forward_implications":["Anatomical landmark detection in brain MRI can be done with one annotated template instead of hundreds of expert-labeled scans, lowering the barrier for new landmark protocols.","A model trained only on T1w data can localize landmarks in T2w scans when contrast augmentation is included, which matters for clinical sites with mixed MRI protocols.","Consistency-based regularization plus registration produces landmarks that follow a pre-defined protocol, unlike registration-driven keypoints that can drift across subjects.","The detected landmarks retain clinical utility: the paper's downstream experiments use inter-landmark distances to separate Parkinson's disease and Alzheimer's disease cohorts with reported F1 scores above 80 percent.","The method tolerates large added rotation misalignments with little degradation, suggesting it can handle the varying head positions common in real acquisitions."],"supporting_citations":[{"why":"Supplies the ICBM152 template used as the single reference image and spatial normalization target.","marker":"[18]"},{"why":"Defines the 32 AFIDs landmark protocol and provides expert-labeled test data used for evaluation.","marker":"[53]"},{"why":"Provides the thin-plate spline deformation model that makes the landmark-to-template warp differentiable.","marker":"[8]"},{"why":"Introduces random convolution as a contrast augmentation technique, which the paper adapts to 3D MRI.","marker":"[65]"},{"why":"Motivates the curriculum-style adaptive loss schedule used to balance registration and consistency.","marker":"[4]"},{"why":"KeyMorph is a keypoint-based self-supervised registration baseline and comparison point for landmark consistency.","marker":"[59]"},{"why":"BrainMorph is a foundational keypoint registration baseline that CABLD is compared against.","marker":"[60]"},{"why":"uniGradICON provides a state-of-the-art registration baseline for comparison.","marker":"[54]"},{"why":"MultiGradICON serves as a multimodal registration baseline in the comparisons.","marker":"[13]"},{"why":"ANTs represents the classical atlas-to-subject registration approach that sets the traditional accuracy bar.","marker":"[3]"}],"fun_headline_variants":["Single template bootstraps brain landmark detection","CABLD: one reference scan finds 32 landmarks","Contrast-agnostic brain landmarks with one example","T1-trained CABLD nails T2 landmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that a thin-plate-spline warp fitted from the model's own initially rough landmark predictions yields a registration loss that is a useful learning signal, so the self-supervised bootstrap converges before the consistency terms take over.","fun_headline_variants_meta":{"raw":{"variants":["Single template bootstraps brain landmark detection","CABLD: one reference scan finds 32 landmarks","Contrast-agnostic brain landmarks with one example","T1-trained CABLD nails T2 landmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000503,"raw_usage":{"total_tokens":2464,"prompt_tokens":958,"completion_tokens":1506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1445}},"tokens_in":574,"tokens_out":1506,"duration_ms":11140,"temperature":1.0,"reasoning_tokens":1445,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:46:57.467757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could retrain CABLD on the same unlabeled T1w scans but replace the two consistency losses with direct regression to the 32 template landmarks on a small labeled subset; if direct supervision produces lower mean radial error, the paper's single-template sufficiency claim is weakened. Alternatively, re-running with the curriculum schedule inverted, putting consistency first and registration second, would test whether the reported convergence and accuracy actually depend on the adaptive ordering.","supporting_citations":[{"cited_title":"Unbiased nonlinear aver- age age-appropriate brain templates from birth to adulthood","cited_arxiv_id":null,"evidence_quote":"Supplies the ICBM152 template used as the single reference image and spatial normalization target."},{"cited_title":"Magnetic resonance imag- ing datasets with anatomical fiducials for quality control and registration","cited_arxiv_id":null,"evidence_quote":"Defines the 32 AFIDs landmark protocol and provides expert-labeled test data used for evaluation."},{"cited_title":"Bookstein","cited_arxiv_id":null,"evidence_quote":"Provides the thin-plate spline deformation model that makes the landmark-to-template warp differentiable."},{"cited_title":"Robust and generalizable visual representation learning via random convolutions","cited_arxiv_id":null,"evidence_quote":"Introduces random convolution as a contrast augmentation technique, which the paper adapts to 3D MRI."},{"cited_title":"Curriculum learning","cited_arxiv_id":null,"evidence_quote":"Motivates the curriculum-style adaptive loss schedule used to balance registration and consistency."},{"cited_title":"A robust and interpretable deep learning frame- work for multi-modal registration via keypoints","cited_arxiv_id":null,"evidence_quote":"KeyMorph is a keypoint-based self-supervised registration baseline and comparison point for landmark consistency."},{"cited_title":"BrainMorph: A Foun- dational Keypoint Model for Robust and Flexible Brain MRI Registration","cited_arxiv_id":null,"evidence_quote":"BrainMorph is a foundational keypoint registration baseline that CABLD is compared against."},{"cited_title":"uniGradICON: A Foun- dation Model for Medical Image Registration","cited_arxiv_id":null,"evidence_quote":"uniGradICON provides a state-of-the-art registration baseline for comparison."},{"cited_title":"MultiGradICON: A Foundation Model for Multimodal Medical Image Registration","cited_arxiv_id":null,"evidence_quote":"MultiGradICON serves as a multimodal registration baseline in the comparisons."},{"cited_title":"Advanced normalization tools (ANTS)","cited_arxiv_id":null,"evidence_quote":"ANTs represents the classical atlas-to-subject registration approach that sets the traditional accuracy bar."}],"review_version":1}