{"id":"4375a317-62bf-4882-8e99-2ea55c143370","arxiv_id":"2412.05580","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised masked mesh network trained on healthy cortical surfaces produces reconstruction-error anomaly scores that separate Alzheimer's patients from controls, but the paper reports no individual-level detection metrics.","lead":"Researchers trained a neural network on healthy brain scans to reconstruct masked parts of the brain's folded outer surface, then used reconstruction errors to flag Alzheimer's-related changes in new people. The method needs no patient labels and points to cortical thickness and other known Alzheimer's biomarkers, though individual-level accuracy is not reported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Age mismatch between training and test populations may drive group-level anomaly differences; without age adjustment or individual-level ROC, the AD-specific claim is not yet established.","rationale":"The reader's weakest assumption matches my concern: the anomaly score lacks validation as an AD-specific signal rather than an age or dataset artifact. The paper's strongest claim depends on this because the only supporting evidence is group-level ROI effect sizes from data with a clear age mismatch and no individual-level metrics. I agree with the CONDITIONAL verdict: the concern is real but testable, and the paper could address it with additional analyses. I find no independent fatal flaw that would require REJECT, and the methodology is plausible; the age/gender embedding is a reasonable design choice but needs empirical confirmation. The reader's call for ROC/AUC, age adjustment, baseline comparison, and code is exactly the right set of checks. Therefore my stress-test does not change the verdict; it reinforces it.","tokens_in":17434,"tokens_out":2236,"duration_ms":23535,"concrete_test":"In ADNI and OASIS3, fit a regression per ROI: anomaly_score ~ age + sex + diagnosis (AD vs HC). If the diagnosis coefficient is not significant after age adjustment in both datasets, the AD-specific claim fails. Additionally, compute the AUC for individual-level AD vs HC classification from the anomaly scores, and repeat in age-matched subgroups (e.g., age difference ≤2 years between AD and HC). If AUC drops to chance in matched subgroups but is high in the full sample, the effect is age-driven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the MMN detects cortical anomalies that are AD biomarkers. The evidence in Section 4.2 is group-level ANOVA effect sizes for ROIs with p<0.05. This evidence is vulnerable to a confound: the model is trained on younger healthy subjects (UKB/HCP-Aging, mean age 56-63 per Table 1) and tested on older ADNI/OASIS3 subjects (mean age 69-75), and within those test sets AD patients are likely older than controls. Age is known to decrease cortical thickness and volume, so the anomaly score defined in Section 3.3 as an L1 reconstruction error may be elevated in AD patients simply because they are older, not because of AD-specific atrophy. The model does include an age/gender embedding (Eq. 2), but the paper never validates that this conditioning removes age effects from the reconstruction error. If the embedding only shifts the predicted mean without capturing age-related variance, or if the training age range does not cover the test age range, the anomaly score will be age-dependent. The absence of individual-level classification metrics (ROC/AUC, detection rates, or age-stratified comparisons) means we cannot distinguish 'detects AD' from 'detects older age' or 'detects dataset shift.' This is the load-bearing weakness: the only reported result consistent with the claim is also fully consistent with an age/dataset artifact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a self-supervised masked mesh learning framework (MMN) for unsupervised anomaly detection on 3D cortical surfaces. The model is trained on healthy subjects from UKB and HCP-Aging to reconstruct masked vertices of cortical feature maps (curvature, sulcus, thickness, volume), using an encoder-decoder with spherical-harmonics-based mesh convolutions and an age/gender conditioning embedding. At test time, anomaly scores are computed per Desikan-Killiany ROI as the L1 reconstruction error when that ROI is masked. The method is evaluated on ADNI and OASIS3 by comparing AD patients to healthy controls using one-way ANOVA with Benjamini-Hochberg correction, and significant ROIs are reported with eta-squared effect sizes. The paper claims that these results demonstrate detection of cortical anomalies that are known biomarkers of Alzheimer's disease.","tokens_in":17685,"tokens_out":4475,"duration_ms":39320,"significance":"If the central claim were fully supported, the MMN would be a valuable contribution: it adapts masked image modeling to cortical meshes, leverages large-scale healthy normative data, and avoids the need for anomaly labels, which is attractive for rare diseases and personalized analyses. The architecture and preprocessing are described in sufficient detail to be reproducible, and the authors state that code will be released. However, the current evaluation is too weak to establish that the framework detects AD-specific anomalies. The group-level ANOVA evidence is vulnerable to age and dataset-shift confounds, no individual-level detection metrics are reported, and no baseline comparisons are provided. The methodological novelty is real, but the empirical support for the stated claims is incomplete.","major_comments":[{"comment":"The training/validation populations are substantially younger (UKB mean age 63.3/63.4, HCP-Aging 56.8/58.0) than the test populations (ADNI 75.5, OASIS3 69.5), and within the test sets AD patients are likely older than controls. Since the anomaly score is defined as an L1 reconstruction error (Section 3.3, Algorithm 1) and cortical thickness and volume are known to decline with age, the observed group differences in anomaly scores could reflect age differences or dataset shift rather than AD pathology. The model includes an age/gender embedding (Eq. 2), but the paper does not validate that this conditioning removes age-related variance from the reconstruction error, nor does it provide age-stratified or age-adjusted analyses (e.g., ANCOVA, age-matched subgroups, or individual-level ROC). This is load-bearing for the central claim that the framework detects AD biomarkers, and the current evidence is equally consistent with an age or dataset artifact.","section":null},{"comment":"The evaluation is limited to group-level ANOVA effect sizes. The abstract and contributions claim that the framework can 'detect anomalies' and 'distinguish dementia from healthy controls,' but no individual-level detection or classification metrics (e.g., ROC-AUC, sensitivity/specificity, detection rate, or a defined anomaly threshold) are reported. The anomaly score is a continuous per-ROI value, yet the paper does not show how this score translates into a decision about an individual being anomalous. Without such metrics, the actual detection performance of the MMN is not established. In addition, no comparison to existing unsupervised anomaly detection baselines is provided, so the practical utility and relative advantage of the method remain unquantified.","section":null},{"comment":"The Euler number quality-control threshold is inconsistent between the text and the figure. Section 4.1.1 states that subjects with a median-centered absolute Euler number greater than 25 were excluded, while Figure 2 (in three places) indicates 'Euler number <=5' as the quality check criterion. This discrepancy affects which subjects are included in the training and test sets and must be resolved for the study to be reproducible and for the reported results to be interpretable.","section":null}],"minor_comments":[{"comment":"The word 'preciese' should be 'precise'.","section":null},{"comment":"The word 'diagnosised' should be 'diagnosed'.","section":null},{"comment":"The figure axis labels read 'Eta' while the text refers to 'Eta Squared (η2)'; the notation should be made consistent.","section":null},{"comment":"The sentence 'These scores were then used to classify subjects into normal and abnormal groups' is misleading because no classification algorithm or threshold is applied; consider rewording to describe the group comparison actually performed.","section":null},{"comment":"The hyperparameter description says values were 'searched according to the original paper,' which is ambiguous; clarify whether values were taken directly from Lei et al. or re-searched on the validation set.","section":null}],"recommendation":"major_revision","confidential_remarks":"The age mismatch between training and test populations is a serious confound that the authors need to address explicitly. If the authors can provide age-adjusted individual-level results (e.g., ROC-AUC on age-matched groups or ANCOVA with age as a covariate) and clarify the Euler number threshold, the paper could be publishable. As it stands, the evidence is too weak to support the abstract's claim of detecting AD biomarkers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a plausible and fairly well-executed application of masked mesh autoencoding to cortical surface anomaly detection, but the evaluation as written does not yet show that the method detects Alzheimer's disease rather than age or dataset shift. It deserves serious review, but the authors need to add individual-level metrics and age control before the central claim holds.\n\nWhat is actually new: the ROI-masked anomaly scoring—masking one Desikan-Killiany ROI at a time and using L1 reconstruction error as a local anomaly score—is a sensible twist on masked image modeling. The evaluation on ADNI and OASIS3 with effect-size maps is more thorough than the prior ISBI paper, and the group-level results in cortical thickness, volume, and sulcus are consistent with known AD atrophy patterns. The large healthy training set from UKB and HCP-Aging is a genuine strength.\n\nSoft spots, in order of severity. First, the age confound is load-bearing. Training data are younger (mean 56–63) than test data (mean 69–75), and within the test sets AD patients are likely older than controls. The model receives age and gender as conditioning inputs, but the paper never validates that this removes age-related variance in reconstruction error. Without age stratification or adjustment, the group differences could be driven by age rather than AD. Second, there are no individual-level detection metrics; only group-level ANOVA effect sizes. ROC or AUC, or detection rates at fixed specificity, would make the practical value clear. Third, no baseline comparison. It is hard to tell what the masked reconstruction adds over simply measuring raw cortical thickness. Fourth, minor issues: the Euler number threshold appears as 25 in the text and 5 in the figure, and code is promised but not released.\n\nI disagree with any claim that this is circular. The anomaly score is not fitted to AD labels, and hyperparameters were tuned on a healthy validation set. The problem is confounded evaluation, not circularity.\n\nThis paper is for researchers in unsupervised anomaly detection in neuroimaging, especially surface-based methods. It is not a breakthrough, but it is a solid step forward if the gaps are addressed.\n\nRecommendation: send it to peer review, with a clear request for major revision: age control, individual-level metrics, a baseline, and code release. If those are fixed, this becomes a useful contribution.","headline":"A plausible masked-mesh anomaly detector with a real age-confound problem; needs individual-level metrics and age control before the AD claim holds.","tokens_in":18246,"tokens_out":3355,"would_cite":false,"duration_ms":30484,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","68T10","92C55","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A masked mesh network trained on healthy brains detects Alzheimer's anomalies without labels.","keywords":["Unsupervised Anomaly Detection","Self-Supervised Learning","Brain Imaging","Cortical Surface","Mesh Convolution","Alzheimer's Disease","Masked Image Modeling","Reconstruction Error"],"falsifier":"Recompute the group comparisons after age-matching AD patients and controls within each test set (or regressing age out of the anomaly scores) and compute individual-level area under the ROC curve for separating patients from age-matched controls; if the effect sizes vanish or the AUC is at chance, the anomaly scores are tracking age or dataset differences rather than Alzheimer's disease.","tokens_in":17236,"feed_emoji":"🧠","tokens_out":7592,"duration_ms":57300,"temperature":0.7,"pith_summary":"The paper aims to show that a self-supervised neural network can learn what normal cortical surfaces look like and then flag an individual's brain as anomalous without any disease labels. The model, a masked mesh convolutional network, is trained on tens of thousands of healthy subjects to reconstruct masked patches of cortical thickness, volume, sulcus, and curvature maps. At test time, it masks each region of a Desikan-Killiany atlas and uses the L1 reconstruction error as an anomaly score. The authors report that these scores separate Alzheimer's patients from controls in two independent test cohorts and highlight cortical regions previously tied to Alzheimer's disease, with thickness the most sensitive feature. If these results hold, the approach offers a label-free, interpretable screening signal for neurodegenerative disease.","feed_headline":"Masked mesh network spots Alzheimer's patterns without labels","feed_subtitle":"A self-supervised model trained on healthy brains flags thickness and volume anomalies in AD patients.","key_machinery":"The central object is the Masked Mesh Network (MMN), a U-Net-like encoder-decoder built from vertex2vertex convolutions whose filters are truncated spherical harmonics, together with mesh pooling and unpooling based on GPU-accelerated mesh decimation. The self-supervised pretext task is masked mesh modeling: half of the vertices are replaced with learnable mask tokens and the model must predict them from visible context, with an age-and-gender embedding concatenated at the bottleneck. This forces the network to encode normal spatial relationships among cortical surface features. For anomaly detection, each Desikan-Killiany ROI is masked in turn, and the L1 distance between the original and reconstructed features in that ROI is the anomaly score, making the score conditional on the rest of the brain and on the subject's phenotype.","core_discovery":"The central claim is that the masked mesh network (MMN), trained solely on healthy cortical surface data, learns a normative model of cortical morphology whose per-ROI reconstruction error acts as an unsupervised anomaly score that detects Alzheimer's disease. The paper demonstrates this at the group level: one-way ANOVAs with Benjamini-Hochberg correction on two independent Alzheimer's disease test sets show significant patient-versus-control differences in many regions, and the effective regions match established AD biomarkers. Cortical thickness yields the largest effect sizes, left-hemisphere regions are more sensitive, and consistent findings across both test sets include superior frontal, precentral, and transverse temporal thickness, fusiform and parahippocampal sulcus, and lateral orbitofrontal volume. The paper frames the ROI-wise masking as a conditional anomaly score with per-subject adaptation, so the same model yields a localized, interpretable map without retraining.","pith_inferences":["Beyond the paper, the decisive missing check is individual-level classification: an area under the ROC curve for AD versus age-matched controls would show whether the group-level effect sizes translate into usable screening accuracy.","Beyond the paper, an age-matched reanalysis within each test set would test whether the embedding actually removes the age confound, since the training and test cohorts differ in mean age by a decade or more.","Beyond the paper, replacing atlas-ROI masking with vertex-wise or sliding-window masking could yield continuous anomaly maps and reveal sub-ROI patterns that the Desikan-Killiany parcellation averages away.","Beyond the paper, the same normative-model logic could extend to other cortical mesh features or other neuropsychiatric conditions, but only after a lifespan reference set is built to avoid confounding development and degeneration."],"forward_implications":["Trained only on healthy subjects' cortical meshes, the MMN yields ROI anomaly scores that separate Alzheimer's patients from controls in two independent datasets.","Cortical thickness is the most sensitive surface feature for AD anomaly detection with this framework, with left-hemisphere regions more sensitive than right-hemisphere ones.","The framework produces interpretable region-level anomaly maps via per-ROI masking, so it localizes disease-relevant cortex without supervised training or voxel-level registration.","The approach does not require large patient cohorts, so the same pipeline could target rare or poorly characterized diseases where collecting labeled patient data is impractical.","The detected anomalous regions overlap with previously reported AD biomarkers, supporting the method's construct validity."],"supporting_citations":[{"why":"Supplies the vertex2vertex convolution and the GPU-accelerated mesh decimation pooling/unpooling blocks that form the MMN architecture.","marker":"Lei et al. (2023)"},{"why":"Provides the masked image modeling pretext task that the paper adapts to cortical meshes.","marker":"Xie et al. (2022)"},{"why":"Gives the U-Net encoder-decoder structure used in the MMN.","marker":"Ronneberger et al. (2015)"},{"why":"Supplies the UKB cohort that forms the bulk of the healthy training and validation data.","marker":"Miller et al. (2016)"},{"why":"Supplies the HCP-Aging cohort used alongside UKB for healthy training and validation.","marker":"Bookheimer et al. (2019)"},{"why":"Provides the first Alzheimer's disease test set with patient labels used for evaluation.","marker":"Petersen et al. (2010)"},{"why":"Provides the second Alzheimer's disease test set with patient labels used for evaluation.","marker":"LaMontagne et al. (2019)"},{"why":"Provides the pipeline that extracts the cortical surface features (thickness, volume, sulcus, curvature) from T1 images.","marker":"Fischl (2012)"}],"fun_headline_variants":["Masked mesh learns brain norms to flag AD anomalies","Unsupervised cortical anomaly detection via masked mesh learning","Self-supervised mesh model detects Alzheimer's patterns","Healthy-brain trained mesh spots AD biomarkers unsupervised","Masked mesh network reveals Alzheimer's anomalies without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the L1 reconstruction error from a model trained on younger healthy adults is a specific signal for Alzheimer's disease rather than a signal for age or dataset artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Masked mesh learns brain norms to flag AD anomalies","Unsupervised cortical anomaly detection via masked mesh learning","Self-supervised mesh model detects Alzheimer's patterns","Healthy-brain trained mesh spots AD biomarkers unsupervised","Masked mesh network reveals Alzheimer's anomalies without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1255,"prompt_tokens":923,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":273}},"tokens_in":539,"tokens_out":332,"duration_ms":3366,"temperature":1.0,"reasoning_tokens":273,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:33:51.246599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the group comparisons after age-matching AD patients and controls within each test set (or regressing age out of the anomaly scores) and compute individual-level area under the ROC curve for separating patients from age-matched controls; if the effect sizes vanish or the AUC is at chance, the anomaly scores are tracking age or dataset differences rather than Alzheimer's disease.","supporting_citations":[{"cited_title":", author Akhtar, N","cited_arxiv_id":null,"evidence_quote":"Supplies the vertex2vertex convolution and the GPU-accelerated mesh decimation pooling/unpooling blocks that form the MMN architecture."},{"cited_title":", author Alfaro-Almagro, F","cited_arxiv_id":null,"evidence_quote":"Supplies the UKB cohort that forms the bulk of the healthy training and validation data."},{"cited_title":", author Aisen, P.S","cited_arxiv_id":null,"evidence_quote":"Provides the first Alzheimer's disease test set with patient labels used for evaluation."},{"cited_title":", year 2012","cited_arxiv_id":null,"evidence_quote":"Provides the pipeline that extracts the cortical surface features (thickness, volume, sulcus, curvature) from T1 images."}],"review_version":1}