{"id":"7a829b1d-c1f8-427f-a9b9-3b66214aa3f8","arxiv_id":"2506.03217","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"petBrain is a fully automated, web-based pipeline that computes Centiloid, CenTauRz, and HAVAs scores for A/T2/N staging, with validation against SPM, B-PIP, fluid biomarkers, and cognition on ADNI.","lead":"This paper introduces petBrain, a web-based pipeline that quantifies Alzheimer's disease biomarkers (amyloid, tau, neurodegeneration) from PET and MRI in about 20 minutes. It is a practical tool that could standardize and speed up biomarker measurement for clinical research.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Calibration transfer is the load-bearing bridge: ADNI validation is pooled over tracers, so the cross-tracer accuracy of the MK/PI and FTM/NAV conversion equations remains unestablished.","rationale":"The reader's weakest-assumption analysis correctly identified calibration generalization as the load-bearing risk. I agree, and I sharpen it with a concrete omission: the ADNI validation is reported as aggregate correlation/ICC across tracers, while the actual sample is dominated by FTP for tau and FBB/FBP for amyloid. The conversion equations for MK/PI (tau) and FTM/NAV (amyloid) are therefore not independently validated by the reported statistics, even though the paper claims cross-tracer comparability. This is not an internal inconsistency; the ADNI comparison is a genuine external check and the high correlations are real evidence. The problem is that the claim of cross-tracer standardization is broader than the evidence presented. A conditional acceptance is appropriate: the pipeline appears valid for the tracers and cohorts tested, but the universal scaling claim needs tracer-stratified validation or an explicitly narrowed scope. I do not see grounds to reject, and the concern is testable without new data collection.","tokens_in":20513,"tokens_out":5724,"duration_ms":60031,"concrete_test":"Stratify the ADNI validation by PET tracer: compute per-tracer R2, ICC, slope/intercept, and Bland-Altman bias for petBrain versus SPM and versus B-PIP, with FBB and FBP separated for CL and FTP, MK, and PI separated for CTRz and tau SUVr, including bootstrapped 95% confidence intervals. If per-tracer bias for MK/PI or FBB/FBP is within the positivity-threshold margins (e.g., about ±5 CL or ±1 CTRz around the cutoff), cross-tracer standardization is supported; otherwise the pooled statistics overstate generalizability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that petBrain produces standardized CL and CTRz values comparable to SPM and B-PIP across tracers via the calibration equations in Eqs. (1)-(2) and Tables 2-3. That claim depends on these affine conversions transferring to all tracers and cohorts. The ADNI validation is strong but pooled over tracers: of 821 tau scans, 780 are FTP, with only 23 MK and 18 PI, and the amyloid validation uses only FBB and FBP, leaving FTM and NAV conversions untested. No tracer-stratified R2/ICC/Bland-Altman results are reported, and no confidence intervals are given for the regression parameters in Tables 2-3. Because the calibration fits were derived from small GAAIN samples (e.g., 79 PiB subjects for the Level-1 CL calibration, 46 FBP subjects for one tracer conversion), tracer- or scanner-specific bias could be masked in the aggregate R2=0.94/0.96 for CL and R2=0.84/0.92 for CTRz. If, for instance, the MK or PI conversion has a systematic offset, T2 status would be misclassified for those tracers despite good pooled performance. The paper's cross-tracer standardization claim is therefore broader than its evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents petBrain, a web-based pipeline that combines AssemblyNet deep-learning segmentation of T1w MRI with PET processing to compute amyloid Centiloid (CL), tau CenTauRz (CTRz), and HAVAs neurodegeneration scores, and thereby to assign A/T2/N status. Calibration equations for CL and CTRz are derived on GAAIN cohorts and validated on ADNI against SPM and B-PIP pipelines. The authors report high concordance (R2 = 0.84 to 0.96, ICC = 0.92 to 0.98), expected associations with CSF and plasma biomarkers, and stronger combined A/T2/N associations with cognition than single-biomarker models.","tokens_in":20794,"tokens_out":7062,"duration_ms":69054,"significance":"If the calibration transfer is reliable, petBrain is a practically valuable contribution: it is freely accessible through the volBrain platform, processes a full A/T2/N workup in about 20 minutes per subject, uses subject-specific deep-learning segmentations, and is the first pipeline described as simultaneously estimating A, T2, and N. The ADNI validation is a genuine external test with strong aggregate concordance against two established pipelines, and the biological correlates (fluid biomarkers, clinical status, cognition) are directionally consistent with the AD literature. The manuscript also ships concrete calibration equations and a public implementation, which aids reproducibility. However, the paper's central cross-tracer standardization claim is broader than the evidence actually presented, and the absence of tracer-stratified validation and uncertainty quantification is load-bearing for that claim.","major_comments":[{"comment":"The ADNI validation is pooled across tracers: 780 of the 821 tau scans are FTP, with only 23 MK and 18 PI scans, and the amyloid validation uses only FBB and FBP, leaving FTM and NAV unvalidated. The conversion equations for MK/PI in Table 3 are derived indirectly from the FTP Level-1 calibration plus published constants, not from petBrain-specific data. Because the central claim is that petBrain produces standardized CTRz values across tracers, the pooled R2=0.84/0.92 could conceal a tracer-specific bias that would directly affect T2 status. Please add tracer-stratified R2, ICC, and Bland-Altman limits of agreement for all tracers represented in the ADNI sample, or explicitly restrict the cross-tracer claim to the tracers that are actually validated.","section":"Section 3.3, Table 1"},{"comment":"The calibration regressions are reported without standard errors, confidence intervals, or residual-bias diagnostics, despite being fitted on small GAAIN samples (for example, 79 PiB subjects for the Level-1 CL calibration and 46 FBP subjects for the FBP conversion). These affine parameters are load-bearing because they map SUVr to CL and CTRz and hence determine the A and T2 binary status. Please report confidence intervals for all slopes and intercepts, and ideally propagate the parameter uncertainty into CL/CTRz values or perform a sensitivity analysis across scanners, tracers, or cohorts.","section":"Tables 2 and 3, Eqs. (1)-(2)"},{"comment":"The HAVAs score is the authors' own previously published model, and the manuscript states that ADNI subjects were excluded from its construction but provides no verifiable documentation or code to support this claim. If any ADNI participants were part of the HAVAs training data, the validation correlations in Tables 4 and 5 for the N biomarker would be partly circular. Please provide the exclusion procedure, a list of overlapping subjects, or a clear description of the HAVAs training cohort separately from the ADNI validation cohort.","section":"Section 2.2, N status and HAVAs"},{"comment":"The subject-specific Centiloid mask is selected on the GAAIN-PiB dataset with a Cohen's d > 5 threshold, and the same dataset is then used for the Level-1 CL calibration in Eq. (1). This double use of the calibration cohort may overstate the reported calibration fit in Supplementary Fig. 3. The ADNI validation is independent, but the stability of the mask-selection procedure itself is not assessed. Please report a bootstrap or resampling-based stability analysis of the selected mask, or demonstrate that the d>5 threshold does not materially change CL values.","section":"Section 2.2, Centiloid subject-specific mask"}],"minor_comments":[{"comment":"The table header lists 'ADNI N = 831', while the text in Section 2.1 states that the study used 821 subjects; the row sums in the table equal 821, so this appears to be a typographical inconsistency that should be corrected.","section":"Table 1"},{"comment":"The caption states that 'Amyloid status was established with the B-PIP PET pipeline and ADNI thresholds,' which conflicts with the surrounding text's presentation of petBrain-derived staging; please clarify which pipeline defines the A-group labels used in the ANOVA analyses in Section 3.4.","section":"Figure 5"},{"comment":"There are inconsistent spellings of 'CenTauR': 'CenTaur' and 'Centaur' appear in Figure 2 and in the main text; the standardized spelling should be used consistently.","section":"Throughout"},{"comment":"The legends section lists 'Figure 2' twice (once for the global overview and once for the comparison with SPM), and the supplementary material contains two different items numbered 'Supplementary Figure 5'; renumbering is needed.","section":"Legends and Supplementary Material"},{"comment":"The text states that 'we used 499 amyloid-PET images' while Table 1 reports GAAIN N=375; the relationship between the number of images, the number of participants, and the tracer-specific counts should be stated more clearly.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The core ADNI validation is convincing and the pipeline is potentially valuable, but the paper's cross-tracer standardization claim exceeds its evidence: the pooled validation masks the near-absence of MK/PI and the complete absence of FTM/NAV validation, and the calibration equations carry no uncertainty. These gaps are fixable within the manuscript's scope through tracer-stratified analyses and confidence intervals, so I recommend major revision rather than rejection. The HAVAs circularity concern also needs explicit documentation of ADNI exclusion from model construction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"petBrain is a genuinely useful engineering contribution: a fast, web-based A/T2/N pipeline that combines Centiloid, CenTauRz, and HAVAs with subject-specific masks, and it validates well on ADNI against SPM and B-PIP. The central quantitative claim holds up in aggregate, but the cross-tracer calibration equations are fitted on small samples and not directly validated per tracer on ADNI, so the standardization claims are broader than the evidence.\n\nWhat's new: the simultaneous A/T2/N output with standardized scales, the ~20-minute runtime via AssemblyNet deep learning segmentation, and the web deployment on volBrain. The ADNI validation is mostly well done—external data, comparisons to two established pipelines, Bland-Altman plots, and expected associations with fluid biomarkers and cognition. The authors also test robustness to PVC and include an honest limitation section (need for MRI, lack of clinician-oriented staging outputs).\n\nThe soft spots are real but not fatal. The biggest is calibration transfer. The Level-1 CL calibration uses 79 PiB scans; the FBP conversion uses 46. For tau, 780 of 821 ADNI scans are FTP, with only 23 MK and 18 PI. The paper reports pooled R2=0.84 for CTRz without tracer-stratified agreement, so a systematic offset in MK or PI could be masked. There are also no confidence intervals on any conversion parameters. The subject-specific CL mask was selected on the same PiB cohort used for calibration, which is mildly circular for that calibration, though the external ADNI validation mitigates it. The 'first integrated A/T2/N pipeline' claim is a bit oversold, since B-PIP already gives A and T2, though not CTRz or N. Code is not released, only the web service; AssemblyNet is open and equations are given, so it is partially reproducible.\n\nOverall, this paper deserves a serious referee. The core result—petBrain's outputs track established pipelines and biology—holds up in aggregate. A referee should push for tracer-stratified validation, confidence intervals on the calibration parameters, and a more measured framing of the 'first' claim. For anyone working on multicenter AD imaging, this is worth citing and trying.","headline":"Solid integrated A/T2/N pipeline with strong aggregate ADNI validation; the cross-tracer calibration claims are broader than the per-tracer evidence.","tokens_in":21398,"tokens_out":2395,"would_cite":true,"duration_ms":24478,"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":"petBrain, a fully automated web pipeline, produces A/T2/N Alzheimer's biomarker scores from PET and MRI that match established SPM and B-PIP pipelines, and it is the first to estimate all three biomarkers simultaneously.","keywords":["A/T/N model","image processing","MRI segmentation","Centiloid","CenTauRz","HAVAs","amyloid PET","tau PET"],"falsifier":"Run petBrain on a cohort scanned with a tracer not in the calibration set (e.g., 18F-RO948 or 18F-GTP1) and compare its CTRz against a reference SPM-based CenTauRz; a systematic residual bias in the linear regression, or an ICC below the 0.92-0.98 range reported on ADNI, would falsify the claim that the affine calibration generalizes.","tokens_in":20304,"feed_emoji":"🧠","tokens_out":10327,"duration_ms":81633,"temperature":0.7,"pith_summary":"This paper introduces petBrain, an automated web-based pipeline that turns a structural MRI plus amyloid-PET and tau-PET scans into the three Alzheimer's disease biomarkers of the A/T2/N framework: Centiloid for amyloid plaques, CenTauRz for tau tangles, and the HAVAs score for neurodegeneration. The authors claim it is the first tool to estimate all three biomarkers simultaneously, using deep-learning segmentation to build subject-specific brain regions in about 15 minutes instead of FreeSurfer's hours. On the ADNI cohort, petBrain's amyloid and tau values agree closely with the established SPM-based and B-PIP pipelines (R2 between 0.84 and 0.96, intraclass correlation between 0.92 and 0.98), and its A/T2/N staging tracks CSF and plasma biomarkers, clinical diagnosis, and cognitive scores. If these results hold, the pipeline would make standardized A/T2/N quantification broadly accessible for clinical research without local software expertise or high-performance computing. The load-bearing step is a set of calibration equations, fitted on small GAAIN datasets, that convert petBrain SUVr values into the universal Centiloid and CenTauRz scales.","feed_headline":"petBrain pipeline reads all three Alzheimer's biomarkers in one run","feed_subtitle":"It matches established SPM and B-PIP pipelines on ADNI, and links A/T2/N scores to blood, CSF, and cognition.","key_machinery":"The machinery is the combination of three components: (1) AssemblyNet, a large ensemble of convolutional neural networks that segments 132 brain structures from a T1-weighted MRI in under 15 minutes, yielding subject-specific regions of interest; (2) affine calibration equations, fitted on GAAIN datasets, that map the partial-volume-corrected SUVr from petBrain's subject-specific masks onto the universal Centiloid and CenTauRz scales; and (3) the HAVAs score, computed from the hippocampal, amygdala, and inferior lateral ventricle volumes against a pathological lifespan model. The subject-specific masks are the enabling object: they replace the fixed standardized masks of the Centiloid and CenTauR projects while the calibration equations preserve cross-tracer comparability, and the same MRI segmentation feeds the N biomarker, which is why a single pipeline can output A, T2, and N together.","core_discovery":"The central discovery is that subject-specific anatomical masks produced by the AssemblyNet deep-learning segmentation can replace the fixed MNI-space masks of the Centiloid and CenTauR projects and still reproduce the same standardized values. After Level-1 calibration with the PiB dataset and Level-2 tracer-to-tracer regressions, petBrain's Centiloid values correlate at R2=0.94 with SPM-based Centiloid and R2=0.96 with B-PIP; CenTauRz values correlate at R2=0.84 with SPM and tau SUVr at R2=0.95 with B-PIP. The pipeline also assigns A+/A-, T2+/T2-, and N+/N- status using external thresholds (AMYPAD/CenTauR/HAVAs), and on ADNI these statuses and scores increase monotonically from amyloid-negative controls through amyloid-positive controls, MCI, and dementia, with combined A/T2/N explaining more variance in cognitive scores than any single biomarker. The authors therefore claim that a single unified, web-deployed pipeline can deliver biologically and clinically meaningful A/T2/N quantification.","pith_inferences":["The calibration-generalization claim is testable outside ADNI: if the same affine equations are fitted per cohort with a bias-correction term, the paper's approach could be extended to novel tracers and scanners, but the authors have not demonstrated that such corrections are needed.","Because PVC changed results only marginally, a future MRI-free or PVC-free variant might preserve A/T2 quantification for research contexts where MRI is unavailable, at the cost of losing the N biomarker.","The high concordance between subject-specific-mask pipelines (petBrain and B-PIP) suggests that deep-learning masks could substitute for FreeSurfer in other quantification pipelines, although the paper does not test this directly.","A natural next step the authors do not take is longitudinal validation: showing that changes in petBrain CL, CTRz, and HAVAs track treatment response or disease progression would determine whether the pipeline is useful for monitoring disease-modifying therapies."],"forward_implications":["Research groups without local imaging software or GPU infrastructure can obtain standardized CL, CTRz, and HAVAs values for a subject in about 20 minutes through the volBrain website.","Because A and T2 are expressed on universal scales, results from different amyloid tracers (PiB, FBP, FBB, FTM, NAV) and tau tracers (FTP, MK, PI, and five more via conversion equations) can be pooled and compared across studies.","The combined A/T2/N model predicts CDR-sb, MMSE, and MoCA better than amyloid load alone, supporting the three-biomarker framework for cognitive staging.","On ADNI, the pipeline's A/T2/N staging separated amyloid-positive CN, MCI, and dementia groups for all three biomarkers with high significance.","The paper claims petBrain is the first pipeline to estimate A, T2, and N simultaneously in one processing framework."],"supporting_citations":[{"why":"Supplies the Centiloid scale, its Level-1 calibration procedure, and the PiB dataset on which petBrain's amyloid calibration is fitted.","marker":"[7]"},{"why":"Supplies the CenTauRz scale, the MetaTemporal mask, the FTP calibration dataset, and conversion equations for other tau tracers.","marker":"[12]"},{"why":"Defines the HAVAs neurodegeneration score and its age-adapted N+/N- threshold.","marker":"[13]"},{"why":"B-PIP, the ADNI consortium pipeline whose published CL and tau SUVr values are the comparison standard for petBrain.","marker":"[15]"},{"why":"AMYPAD consensus recommendations that set the CL thresholds for A-, intermediate, and A+ status in the online pipeline.","marker":"[18]"},{"why":"AssemblyNet, the deep-learning segmentation ensemble that produces the subject-specific ROIs underlying all petBrain measurements.","marker":"[19]"},{"why":"Provides the 24.1 CL threshold used in the experiments for A+/A- classification.","marker":"[30]"},{"why":"SPM25, the SPM-based reference pipeline against which petBrain's CL and CTRz values were validated.","marker":"[31]"}],"fun_headline_variants":["Deep-learning pipeline nails all three Alzheimer's biomarkers in one shot","Subject-specific masks let one pipeline read amyloid, tau, and atrophy","One web pipeline matches standard tools for Alzheimer's PET and MRI","petBrain's deep-learning masks match gold-standard Alzheimer's reads","Unified pipeline quantifies amyloid, tau, and neurodegeneration from PET and MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The calibration-generalization assumption: the linear equations that convert petBrain SUVr to Centiloid and CenTauRz scales were fitted on small GAAIN cohorts and are assumed to hold across scanners, tracers, and populations without residual bias.","fun_headline_variants_meta":{"raw":{"variants":["Deep-learning pipeline nails all three Alzheimer's biomarkers in one shot","Subject-specific masks let one pipeline read amyloid, tau, and atrophy","One web pipeline matches standard tools for Alzheimer's PET and MRI","petBrain's deep-learning masks match gold-standard Alzheimer's reads","Unified pipeline quantifies amyloid, tau, and neurodegeneration from PET and MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000871,"raw_usage":{"total_tokens":3800,"prompt_tokens":1004,"completion_tokens":2796,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2705}},"tokens_in":620,"tokens_out":2796,"duration_ms":18730,"temperature":1.0,"reasoning_tokens":2705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:21:34.012496+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run petBrain on a cohort scanned with a tracer not in the calibration set (e.g., 18F-RO948 or 18F-GTP1) and compare its CTRz against a reference SPM-based CenTauRz; a systematic residual bias in the linear regression, or an ICC below the 0.92-0.98 range reported on ADNI, would falsify the claim that the affine calibration generalizes.","supporting_citations":[{"cited_title":"As illustrated in Supplementary Figure 4, the correlation between petBrain and B-PIP remained highly consistent whether PVC was applied (see Figure 3) or not","cited_arxiv_id":null,"evidence_quote":"Supplies the Centiloid scale, its Level-1 calibration procedure, and the PiB dataset on which petBrain's amyloid calibration is fitted."},{"cited_title":"Supplementary Table 3: Cohen’s d scores for each structure on the young CN A- and old AD A+ of the FTP dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the CenTauRz scale, the MetaTemporal mask, the FTP calibration dataset, and conversion equations for other tau tracers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"B-PIP, the ADNI consortium pipeline whose published CL and tau SUVr values are the comparison standard for petBrain."}],"review_version":1}