{"id":"374f2be3-3856-426c-8ea6-de04b582b7ce","arxiv_id":"2412.01865","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding AI-synthesized cerebral blood volume maps to T1w MRI improved brain age prediction MAE from 4.10 to 3.95 years on a 288-scan test set.","lead":"Researchers combined standard structural brain scans with AI-generated blood-flow maps to predict brain age, reporting a small accuracy gain over structural scans alone. The method could make a vascular-informed brain-age biomarker available without injecting contrast agents, if the synthetic blood-flow maps truly add independent information.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that AICBV adds functional information is unsupported because AICBV maps are synthesized from the same T1w scans, so the multimodal gain may reflect ensembling rather than independent vascular measurement.","rationale":"The reader's weakest_assumption matches my own: the multimodal claim depends on AICBV carrying functional information not already present in T1w. The paper's own methods show AICBV is synthesized from T1w, so the two modalities are not independent measurements. This is a fundamental conceptual issue, not merely a missing validation. The performance gain in Table 1 is small (0.14 years MAE), no confidence intervals are reported, and the ANOVA is unspecified. The abstract's MCI/AD concordance claims, including p=1.47e-8, are absent from the body, further weakening the clinical interpretation. Given these problems, the reader's REJECT verdict is appropriate, and my stress-test does not change it. The concrete test I propose would settle whether the gain is due to functional content or to generic ensembling; if the placebo ensemble reproduces the improvement, the central claim fails. I found no reason to disagree with the reader's assessment, and I have not identified a different load-bearing concern that would alter the verdict.","tokens_in":11311,"tokens_out":3269,"duration_ms":31692,"concrete_test":"Run the identical training pipeline, with identical splits and linear regression, replacing AICBV with a placebo synthetic modality derived deterministically from each T1w scan, e.g., a voxelwise local gradient magnitude map or a second randomly initialized VGG's penultimate feature maps, neither of which contains vascular information. If the T1w + placebo ensemble achieves the same test MAE/R2 improvement over the T-model as the TA-model within bootstrap confidence intervals, then the claimed functional independence is refuted and the gain is attributable to ensembling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that AICBV maps provide functional/vascular information complementary to T1w MRI. However, Section 2.2 states that AICBV is generated from each T1w MRI scan using a 3D patch-based CNN-Mamba model. At inference, AICBV is therefore a deterministic function of the T1w input: AICBV = g(T1w). Any information contained in AICBV is a transformation of the same structural scan, not an independently measured functional modality. The reported improvement of the TA-model over the T-model (MAE 4.10 vs. 3.96, Table 1) could be explained by ensembling two networks trained on the same underlying image, which typically reduces prediction variance, rather than by the addition of true vascular information. Section 4.1 calls AICBV 'an additional independent covariate,' but the paper does not demonstrate information-theoretic or empirical independence. The Grad-CAM results in Section 3.4 show different regions of focus for the two encoders, but different learned features of the same input do not establish that AICBV captures variance in brain age beyond T1w. The AICBV generator from prior work [12] is not released and is not revalidated on the 13 datasets used here, so the core premise is unverified. A control experiment with a synthetic non-vascular modality generated from T1w would directly test whether the observed gain is specific to AICBV or generic to ensembling.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a brain-age estimation pipeline in which two separate 3D VGG-based regressors are trained, one on T1-weighted MRI and one on AICBV maps synthesized from those same T1w scans by a CNN-Mamba generator, and the predicted ages are then combined by linear regression, optionally with sex as an additional covariate. On a single split of 2,851 scans from 13 public datasets, the combined model achieves MAE 3.96 and R² 0.943, compared with MAE 4.10 for the T1w-only model. The abstract additionally reports clinical concordance results (stable vs. progressive MCI separation, correlations with CDRSB and MMSE) that do not appear in the body of the manuscript.","tokens_in":11622,"tokens_out":5076,"duration_ms":50329,"significance":"If the reported improvement is robust and the AICBV maps truly carried independent vascular information, the approach would be a clinically attractive, contrast-free way to augment brain-age biomarkers with functional information. The paper has concrete strengths: it uses a large multi-site sample of 13 public datasets, a full 3D architecture, an interpretability analysis with Grad-CAM, and it directly compares with a prior T1w-only model. However, the central premise is not established: AICBV is a deterministic function of the same T1w input, so the observed gain may reflect ensembling of two views of one structural scan rather than structural-plus-functional fusion. The abstract also makes clinical claims that cannot be verified from the body. The significance of the work therefore depends on additional validation and a major reframing of the claims.","major_comments":[{"comment":"The core claim that AICBV provides an 'additional independent covariate' is not supported. As stated in §2.2, AICBV is generated from each T1w MRI scan by a 3D patch-based CNN-Mamba model, so at inference AICBV = g(T1w) is a deterministic function of the same structural input. Any information in AICBV is therefore a transformation of the T1w scan, not an independently measured vascular modality. The improvement of the TA-model over the T-model (MAE 3.96 vs. 4.10) could be a variance-reduction/ensembling effect from combining two networks trained on different transforms of the same image. To support the multimodal claim, the authors should provide a control experiment using a synthetic non-vascular modality derived from T1w (e.g., a smoothed or noise-transformed T1w) and test whether the TA gain is specific to AICBV; they should also report residual correlations and prediction-error correlations between the T-model and A-model, and, if possible, validate AICBV against actual contrast-enhanced CBV on at least a subset of the 13 datasets. The AICBV generator from reference [12] is not released, which further prevents readers from assessing what the AICBV maps contain.","section":"§2.2, §4.1"},{"comment":"The abstract reports clinical concordance results that are absent from the Methods, Results, and Discussion sections: the separation between stable and progressive MCI with p = 1.47×10⁻⁸, the CDRSB correlation r = 0.403, and the MMSE correlation r = -0.310. None of these analyses appears in the body of the manuscript, and no table or figure reports them. This is a load-bearing omission because these results are used to argue clinical utility. The authors must either add the full analysis (including subject selection, statistical procedure, and appropriate multiple-comparison handling) or remove these claims from the abstract.","section":"Abstract vs. Body"},{"comment":"The ANOVA procedure is not described, and the reported p-values are not interpretable without details. The text says 'we performed ANOVA statistical tests' but does not state what the response variable was, how the model comparison was structured, whether the test was on per-subject absolute errors, whether it was repeated over multiple data splits, or how the p-values were adjusted. In addition, all results are based on a single 8:1:1 split with no confidence intervals for MAE, MSE, or R², and no measurement of split-to-split variability. Given that the absolute improvement of the TA-model over the T-model is only 0.14 years, the authors should demonstrate that this difference is stable across repeated splits and report effect sizes and confidence intervals.","section":"§3.1, Table 1"},{"comment":"The Grad-CAM analysis does not establish that AICBV contributes independent functional information. The observation that the T1w and AICBV encoders focus on different regions is expected when the two encoders are trained on different transformations of the same underlying scan; different learned features of the same input do not demonstrate that AICBV captures variance in brain age beyond T1w. The authors should either connect the saliency maps to an external validation of AICBV, or temper the interpretation that the maps reveal 'functional features that purely anatomical information from T1w MRI cannot provide.'","section":"§3.4"},{"comment":"The interpretation of the sex coefficient appears inconsistent with the coding described. The text states that sex was coded as 0 for female and 1 for male and that the coefficient was -0.168. The authors then write that this 'suggests that the model adjusts male age predictions downward, indicating that male brains are perceived as appearing older than female brains.' A negative coefficient for male would lower predicted age for males, implying that male brains are predicted as younger, not older. This should be corrected or clarified, because the directional claim is part of the validation narrative.","section":"§4.2"}],"minor_comments":[{"comment":"Table 3 uses 'DLBS' while the text uses 'dlbs'; Tables 2 and 3 label the combined model with sex as 'TAG-model' whereas the text and Table 1 use 'TAS-model.' Please unify all abbreviations.","section":"§3.3, Tables 2 and 3"},{"comment":"The text refers to the AICBV generation method as 'the DeepC algorithm,' but §2.2 and reference [12] describe a 3D patch-based CNN-Mamba model. The name 'DeepC' is not introduced or defined, which is confusing.","section":"§4.1"},{"comment":"The future-work paragraph promises 'unsupervised learning' and 'large language models' without explaining how these would integrate with the current regression framework; consider either trimming or making these directions concrete.","section":"§5, Future Work"},{"comment":"There are minor typographical errors (e.g., 'occuring' in the Introduction, 'andR2' in §3.1) and a few sentences with missing spaces. A careful proofreading pass is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The empirical comparison is a useful starting point, but the manuscript overstates the multimodality claim and the abstract contains results that are not in the body. If the authors cannot provide a control experiment or external validation of AICBV, the paper may not be salvageable in its current form; however, the issues are addressable in principle through additional analyses and a revised framing, so I recommend major revision rather than immediate rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a straightforward extension of the group's own AICBV method, applied to brain age. The work is clearly presented and the multi-dataset training (13 datasets, 2,851 scans) is a plus. The late-fusion VGG8 + linear regression is standard but sensible. On the test set, the combined T1w + AICBV model reaches MAE 3.96 vs 4.10 T1w-only—a small but reproducible-sounding gain. The Grad-CAM analysis shows the two encoders attend to different regions, which is a nice exploratory step.\n\nThe problems start when the abstract claims more than the body delivers. The abstract states that AICBV-based BrainAGE separates stable vs progressive MCI with p=1.47e-8 and gives correlations with CDRSB and MMSE; none of these analyses appear in the paper. That is a serious reporting gap. The ANOVA described in Table 1 is not specified (paired? repeated measures? on which residuals?), and no confidence intervals or split variability are reported, so the 0.14-year MAE improvement could easily be noise.\n\nMore substantively, the paper calls AICBV an 'additional independent covariate.' But AICBV is generated from each T1w scan by a learned model (Section 2.2). At inference, AICBV = g(T1w), so it carries no information that is not already a transform of the structural image. The gain from adding AICBV-predicted age to T1w-predicted age may simply reflect ensembling two networks trained on the same input, not the addition of true vascular information. The paper does not discuss this. A control experiment—for instance, fusing predictions from a second T1w-trained network with different initialization, or using a synthetic non-vascular modality derived from T1w—would directly test whether the gain is specific to AICBV or generic to ensembling. That experiment is missing.\n\nThe AICBV generator itself is not released and is not revalidated on these 13 datasets, so we are taking the prior paper's word that the maps represent cerebral blood volume. That may be acceptable for an incremental study, but it does not justify calling the modality 'functional' here.\n\nMy overall read: the work is honest in its core numbers, clearly written, and the question—could synthetic perfusion-like maps help brain age?—is a reasonable one. But as presented, the evidence does not back the abstract's clinical claims, and the interpretation of 'multimodal' is not established. A serious referee could turn this around with requested revisions: put the clinical analyses in the body, give CIs and a proper description of the ANOVA, and add the control experiment. I would not cite this in its current form. I would send it to peer review, though, because the methodological question is legitimate and the paper is well enough constructed to be fixable.","headline":"The empirical gain is small and real, but the paper overclaims clinical results missing from the body, and the 'functional' addition may be just ensembling.","tokens_in":12208,"tokens_out":2957,"would_cite":false,"duration_ms":26848,"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":"Adding AI-synthesized blood-volume maps to structural MRI cuts brain-age error to 3.96 years.","keywords":["brain age gap estimation","BrainAGE","AICBV","cerebral blood volume synthesis","3D convolutional neural network","multimodal MRI","T1-weighted MRI","mild cognitive impairment"],"falsifier":"On held-out subjects with both T1w MRI and true contrast-enhanced steady-state CBV maps (gadolinium or ferumoxytol), compute voxel-wise correlation between AICBV and true CBV in periventricular and dentate-gyrus regions; if that correlation is no higher than the correlation between true CBV and a nonlinear intensity transform of T1w, the AICBV maps carry no independent vascular signal and the multimodal interpretation fails.","tokens_in":11096,"feed_emoji":"🧠","tokens_out":13713,"duration_ms":113512,"temperature":0.7,"pith_summary":"The paper is trying to establish that brain-age estimation—usually built from structural T1-weighted MRI alone—becomes both more accurate and more clinically informative when AI-synthesized cerebral blood volume (AICBV) maps, generated from the same T1w scan, are added as a second modality. It reports that a late fusion of two 3D VGG-based age predictors reaches a mean absolute error of 3.96 years and $R^2$ = 0.943 on held-out scans, beating T1w-only (4.10) and AICBV-only (4.49) models. It also reports that the AICBV branch highlights vascular-rich and periventricular regions, especially in older adults, and that AICBV-based BrainAGE separates stable from progressive mild cognitive impairment. If the claim holds, routine non-contrast MRI could produce a functional-vascular brain-aging biomarker without extra scanning or contrast agents.","feed_headline":"Brain-age error drops to 3.96 years with synthetic blood-volume maps","feed_subtitle":"Combining structural T1w MRI with AI-synthesized cerebral blood volume beats T1-only models.","key_machinery":"The carrying mechanism is a two-stream late-fusion pipeline. A 3D patch-based CNN-Mamba generator, the AICBV model from reference [12], synthesizes a cerebral blood volume map from each non-contrast T1w scan; two separate 3D VGG8 networks then predict age, one from T1w and one from AICBV; and a linear regression combines the two predicted ages, optionally with sex, into the final brain age estimate. The load-bearing element is the AICBV map itself, since the claimed improvement depends on it carrying functional blood-volume information that is not already visible in the T1w input. Grad-CAM heatmaps from the final convolutional layer of each encoder show which regions drive each modality's prediction.","core_discovery":"The authors' central discovery claim is that a multimodal BrainAGE model formed by linearly regressing the age predictions of a T1w-trained 3D VGG8 network and an AICBV-trained 3D VGG8 network outperforms either modality alone: MAE 3.96 years and $R^2$ = 0.943 on the 288-subject test set, versus MAE 4.10 for T1w-only and 4.49 for AICBV-only, with sex as a covariate yielding 3.95 years. They interpret the gain as evidence that AICBV supplies independent functional vascular information absent from T1w, supporting this with Grad-CAM saliency maps that show the T1w encoder emphasizing white-matter and cortical atrophy while the AICBV encoder emphasizes central vascular-rich and periventricular regions, shifting from medial prefrontal cortex in younger adults to the dentate gyrus in older adults. The abstract further claims that BrainAGE values increase across cognitively normal, MCI, and Alzheimer's strata and that AICBV-based BrainAGE separates stable from progressive MCI at $p = 1.47 \\times 10^{-8}$.","pith_inferences":["A natural null control would feed the same late-fusion design with two views of the T1w scan, e.g., an autoencoder's reconstruction as the 'second modality'; if accuracy matches the TA-model, the AICBV branch may be capturing structure rather than independent function.","If AICBV carries true vascular signal, the method should transfer to other structural contrasts or external cohorts not among the 13 training sets, with a gain proportional to how much functional information the structural input lacks.","The abstract's stable-vs-progressive MCI result is stated without a corresponding detailed analysis in the supplied body; reproducing that analysis on ADNI would be the natural check before relying on it.","Because AICBV is generated deterministically from T1w, the two branches are not independent measurements; comparing AICBV against true contrast-enhanced CBV would settle how much functional signal this approach actually adds."],"forward_implications":["Brain-age estimation can extract functional-vascular information from existing non-contrast MRI scans, eliminating the need for contrast agents or additional perfusion scans.","The reported MCI-separation result implies that AICBV-based BrainAGE could serve as an early, non-invasive marker of prodromal cognitive decline, complementing atrophy-based biomarkers.","The age-stratified results imply that structural and vascular modalities contribute differently across the lifespan, with AICBV carrying more weight in older age groups and T1w in younger groups.","The small but significant sex coefficient (MAE 3.95, $p$ = 0.02) is consistent with the model encoding known sex differences in brain aging.","Grad-CAM's focus on periventricular regions and the dentate gyrus suggests the model tracks known aging-vulnerable vascular territories, which could make predictions easier for clinicians to interpret."],"supporting_citations":[{"why":"Supplies the AICBV maps: this is the pretrained 3D patch-based CNN-Mamba generator that turns non-contrast T1w scans into synthetic cerebral blood volume maps, the functional modality at the heart of the multimodal claim.","marker":"[12]"},{"why":"Defines the T1w-only baseline and architecture: a VGG-based BrainAGE model on the same heterogeneous dataset with MAE 4.06 and $R^2$ 0.941, the benchmark the combined model claims to surpass.","marker":"[6]"},{"why":"Establishes prior evidence that synthesized functional MRI from structural scans carries clinical signal for Alzheimer's disease, motivating the use of AICBV-like synthetic maps for brain-age prediction.","marker":"[23]"},{"why":"Identifies aging-vulnerable brain regions used to interpret the complementary Grad-CAM saliency patterns of the T1w and AICBV encoders.","marker":"[5]"}],"fun_headline_variants":["Multimodal brain age: T1 plus synthetic blood volume beats T1 alone","Combining MRI and AI-synthesized blood volume sharpens brain age estimates","Synthetic cerebral blood volume cuts brain-age error to 3.95 years","Vascular maps improve brain age gap detection in aging and MCI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that AICBV maps synthesized from T1w MRI carry functional vascular information that is not already contained in the T1w structural scan; if AICBV is only a learned nonlinear re-expression of T1w tissue contrast, the multimodal gain is an ensemble effect of two views of the same structural image, not structural-plus-functional fusion.","fun_headline_variants_meta":{"raw":{"variants":["Multimodal brain age: T1 plus synthetic blood volume beats T1 alone","Combining MRI and AI-synthesized blood volume sharpens brain age estimates","Synthetic cerebral blood volume cuts brain-age error to 3.95 years","Vascular maps improve brain age gap detection in aging and MCI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3277,"prompt_tokens":1178,"completion_tokens":2099,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":794,"completion_tokens_details":{"reasoning_tokens":2017}},"tokens_in":794,"tokens_out":2099,"duration_ms":12527,"temperature":1.0,"reasoning_tokens":2017,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:48:20.902053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On held-out subjects with both T1w MRI and true contrast-enhanced steady-state CBV maps (gadolinium or ferumoxytol), compute voxel-wise correlation between AICBV and true CBV in periventricular and dentate-gyrus regions; if that correlation is no higher than the correlation between true CBV and a nonlinear intensity transform of T1w, the AICBV maps carry no independent vascular signal and the multimodal interpretation fails.","supporting_citations":[{"cited_title":"Jomsky et al","cited_arxiv_id":null,"evidence_quote":"Supplies the AICBV maps: this is the pretrained 3D patch-based CNN-Mamba generator that turns non-contrast T1w scans into synthetic cerebral blood volume maps, the functional modality at the heart of the multimodal claim."},{"cited_title":"Neurobiology of Aging 91, 15–25 (Jul 2020)","cited_arxiv_id":null,"evidence_quote":"Defines the T1w-only baseline and architecture: a VGG-based BrainAGE model on the same heterogeneous dataset with MAE 4.06 and $R^2$ 0.941, the benchmark the combined model claims to surpass."},{"cited_title":"PLoS One 15(7), e0234255 (Jul 2020)","cited_arxiv_id":null,"evidence_quote":"Identifies aging-vulnerable brain regions used to interpret the complementary Grad-CAM saliency patterns of the T1w and AICBV encoders."}],"review_version":1}