{"id":"3962adea-ed5e-45ac-92f5-f01084964a19","arxiv_id":"2411.13490","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"SKAD, a Sobel-gradient angle difference feature computed after MRI registration, matches Jacobian-map accuracy for dementia classification while cutting the post-registration preprocessing time by roughly 84%.","lead":"This paper proposes a fast image preprocessing step, called SKAD, that uses Sobel edge-detection kernels to create brain maps for detecting Alzheimer's disease and dementia from MRI scans. The authors report that SKAD is about six times faster than the standard Jacobian-map method and reaches similar classification accuracy on two public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never establishes that the Eq. 7 SKAD angle measures volumetric change; it is an intensity-gradient orientation difference, so the claimed equivalence to Jacobian maps and the 6.3x efficiency argument rest on an unvalidated proxy.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: Eq. 7 measures an angle between image-intensity gradients, while a Jacobian map measures a deformation-derived volume ratio. The paper asserts, but does not derive or validate, that these are interchangeable. My read of the manuscript confirms this gap: Section III-C motivates SKAD by convolution stability but never shows that the angle difference is related to det(J), and Figure 3 is illustrative rather than quantitative. The computational savings in Table I and Fig. 5 are credible, and the ablation shows SKAD performs comparably in accuracy; however, if SKAD does not actually encode volumetric change, the method is a different feature extractor, not a Jacobian-map alternative. The proposed synthetic-deformation test would settle this directly, and voxelwise agreement on real data would address generalizability. Because the concern is real but testable, the prior CONDITIONAL verdict remains appropriate; no change to the verdict is needed, but the condition should explicitly include validation of the volume-change claim.","tokens_in":11518,"tokens_out":3193,"duration_ms":38004,"concrete_test":"On ADNI or OASIS, compute both the Jacobian determinant map and the SKAD map for the same subjects and measure voxelwise agreement (e.g., Spearman correlation and overlap of top atrophy voxels). Then run a controlled synthetic check: deform the MNI template with diffeomorphic fields whose Jacobian determinant is known exactly, compute SKAD from the warped image and template, and measure the rank correlation between SKAD and the true det(J). If the correlation is weak or depends on image contrast, SKAD cannot be claimed to quantify volume change.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SKAD (Eq. 7) 'captures regional volume variations' and is a valid substitute for Jacobian maps. A Jacobian map is the determinant of the deformation gradient obtained from nonlinear registration; it measures local volume change independently of image intensity. SKAD instead convolves the warped patient image and the fixed template with Sobel kernels and takes the absolute difference of the 3D angle formed by their intensity gradients. No derivation or empirical evidence connects this angle to det(J). Under an intensity-preserving diffeomorphic warp M∘φ, the gradient transforms as ∇(M∘φ) = Dφ^T (∇M)∘φ, so the angle depends on the full local deformation gradient and on the local intensity structure, not on volume change alone; in flat regions the angle is undefined (Eq. 7 also divides by ∇Mz). The assertion in Sec. III-C that the angle 'signifies abnormalities or atrophy' is therefore unsupported. Since the efficiency argument is premised on SKAD being an alternative to Jacobian maps, the load-bearing step is this missing validation, not the timing comparison, which is credible on its own.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Sobel Kernel Angle Difference (SKAD) as a computationally cheaper alternative to Jacobian maps for quantifying brain volume changes in dementia classification. SKAD is computed by convolving the registered patient MRI and the MNI template with 3D Sobel kernels, then taking the absolute difference of the polar angle formed by the three gradient components (Eq. 7). The authors evaluate SKAD versus registered images and Jacobian maps as inputs to a fine-tuned VGG-16 on ADNI and OASIS-3, report time and FLOP savings (93 ms vs 584 ms, 1.66B vs 146B FLOPs), and compare against prior Jacobian-based works.","tokens_in":11820,"tokens_out":2657,"duration_ms":31639,"significance":"If the central claim were established, the efficiency gain (84% time reduction, 99% FLOP reduction) would be practically useful for large-scale neuroimaging pipelines. The paper is credit-worthy in several respects: the Sobel kernels and angle formula are fixed a priori with no fitted constants, the within-paper ablation across registered, Jacobian, and SKAD inputs is honest, the subject-disjoint split is explicitly stated, and the profiling numbers are concrete and falsifiable. However, the conceptual foundation of the method — the claim that Eq. 7 captures regional volume variations — is not validated, and the empirical evaluation lacks statistical rigor. The significance of the work therefore depends on a load-bearing assumption that the manuscript does not yet support.","major_comments":[{"comment":"The central claim that the angle difference between Sobel gradients of the warped image and the template signifies atrophy or volume change is unsupported. For an intensity-preserving warp, ∇(M∘φ) = Dφ^T (∇M)∘φ, so the angle depends on the full local deformation gradient and on the local intensity structure, not solely on the Jacobian determinant. In anatomically uniform regions the gradient magnitude can approach zero, making the ratio (√(∇Mx²+∇My²))/∇Mz unstable, and the absolute difference of arctangents does not handle the full 3D orientation (it reduces to the polar angle from the z-axis). No derivation, no voxel-wise comparison with det(J), and no empirical correlation is provided. Since the efficiency argument rests on SKAD being a substitute for Jacobian maps, please add a mathematical or empirical validation, for example a scatter plot or correlation coefficient between SKAD values and det(J) on a held-out set.","section":"Section III-C, Eq. (7)"},{"comment":"The empirical evaluation does not report error bars, confidence intervals, or statistical tests. Table II compares SKAD with prior works that use different models (CNN, RF+CNN), different numbers of input modalities (MRI only, MRI+CT, MRI+clinical data), and different datasets (ADNI vs OASIS-3), so the reported accuracies are not matched comparisons. The within-paper ablation in Table III is more informative, but without repeated runs or cross-validation folds it is impossible to know whether the 1–3 point gaps between Jacobian and SKAD are meaningful. Please report mean ± standard deviation over at least 5 runs or cross-validation folds, and restrict claims of competitiveness to the controlled within-paper comparison.","section":"Section IV, Tables II and III"},{"comment":"The data split and augmentation protocol are incompletely specified. The rebalancing to (1233, 1233, 1233, 1233) for OASIS and (133, 133, 133, 133) for ADNI is described for the training set, but it is not stated whether the test set is augmented, whether the reported metrics are on the original test set, or how many unique subjects are in each split. Since augmentation can leak into test performance if applied after splitting, please state explicitly that augmentation is applied only to training images after the subject-disjoint split, and report the number of subjects (not scans) per class in train and test sets.","section":"Section IV-A"}],"minor_comments":[{"comment":"The phrase \"captures regional volume variations\" is used before the connection to volume change is established; consider rewording to \"gradient orientation disparity\" pending the validation requested above.","section":"Abstract and Section III-C"},{"comment":"The 6.3x speedup compares SKAD (93 ms) with the Jacobian map (584 ms) in isolation, but the full pipeline still requires the 12,972 ms registration step. Please clarify in the text that the speedup applies only to the post-registration volumetric-change computation, not to the end-to-end pipeline.","section":"Section IV-C, Table I"},{"comment":"The claim that SKAD \"demonstrates the most stability\" is based on visual inspection of training curves; reporting variance across runs would make this quantitative.","section":"Figure 6"},{"comment":"Minor typographical issues include \"ADIN\" in the Figure 6 caption, \"NVDIA\" in Section IV-C, and a missing period between two sentences in the abstract.","section":"Throughout"},{"comment":"Reference [6] appears as \"V oxel-based morphometry\" and should be corrected to \"Voxel-based morphometry.\"","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is best understood as a methods/empirical paper proposing SKAD as an efficient input representation for dementia classification. The main barrier is not the absence of a theoretical derivation per se, but the absence of any validation that the proposed quantity relates to volume change. If the authors add a correlation analysis against Jacobian determinants and proper statistical reporting, the paper could become a useful contribution. The current comparison with prior work in Table II is too loose and should be reframed. I also note that the journal scope should consider whether the efficiency claim (6.3x) is appropriately contextualized, since registration still dominates the pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this paper proposes SKAD, a fixed 3D Sobel angle-difference map between the registered MRI and the MNI template, as a cheap input feature for dementia classification. The components are standard (3D Sobel kernels, gradient-angle formula), and this specific combination for dementia classification is new as far as the cited literature goes. The paper does not derive new theory; it evaluates an engineering choice.\n\nThe paper does some things well. The ablation is honest: the same VGG-16 setup is compared across registered images, Jacobian maps, and SKAD on two datasets, and SKAD lands close to Jacobian in accuracy while being clearly cheaper (93 ms vs 584 ms, 1.66B vs 146B FLOPs) for the post-registration step. The authors also fairly note that registration itself dominates the pipeline (12,972 ms), so the speed gain is real but scoped. SKAD has no hidden fitted constants; it is reproducible from the equations.\n\nThe soft spots are in the conceptual claim and the evaluation rigor. Equation 7 is presented as capturing 'regional volume variations,' but the angle between intensity gradients of the warped image and the template is not the Jacobian determinant. Under a warp, the gradient transforms by the full deformation gradient, so the angle depends on intensity structure, not volume change alone; in flat regions the formula is undefined. The paper asserts the link but never derives or empirically validates it. That is fixable: either correlate SKAD with Jacobian maps on a validation set or soften the language to 'a discriminative structural feature.' The evaluation also lacks error bars and statistical tests; the minority classes are augmented heavily (e.g., OASIS SEV from 20 to 1,233), which can inflate metrics; Table II mixes models and datasets, making the state-of-the-art comparison hard to read; and no code or data are released. These are all addressable.\n\nThis paper is for researchers working on cheap preprocessing for neuroimaging classification. It deserves a serious referee: the core idea is checkable, the ablation is a useful data point, and the speed advantage is concrete. But I would expect major revision, not acceptance as is. Send it to review with a clear request to validate or reframe the volume-change claim, add statistics, and release code.\n\nRecommendation: engage with it; don't desk reject.","headline":"SKAD is a cheap, plausible input feature for dementia MRI, but the paper overclaims that an intensity-gradient angle measures volume change; the empirical case is promising but needs statistical rigor and validation.","tokens_in":12256,"tokens_out":3944,"would_cite":false,"duration_ms":38688,"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":"The paper claims that Sobel gradient angle differences (SKAD) match Jacobian maps for dementia detection while cutting per-image computation by about 6.3x.","keywords":["Alzheimer's disease","dementia","Jacobian maps","Sobel kernel","voxel-based morphometry","structural MRI","convolution","atrophy detection"],"falsifier":"Compute both the SKAD map and the Jacobian determinant map from the same set of registered MRI scans and measure the voxelwise correlation between SKAD values and log-Jacobian determinants. If the correlation is near zero, or if SKAD's classification performance comes primarily from template registration artifacts rather than from true volume change, then the claim that SKAD captures regional volume variations is falsified.","tokens_in":11253,"feed_emoji":"🧠","tokens_out":6275,"duration_ms":58579,"temperature":0.7,"pith_summary":"The paper proposes Sobel Kernel Angle Difference (SKAD) as a cheap stand-in for Jacobian maps in voxel-based morphometry. SKAD computes the angle between 3D Sobel gradients of a registered patient MRI and the MNI template, and uses that angle difference as a signal of regional brain-volume change. The authors report that SKAD runs in 93 ms per image versus 584 ms for a Jacobian map (about 6.3x faster, and roughly 88x fewer floating-point operations), while a VGG-16 classifier trained on SKAD inputs reaches comparable accuracy to one trained on Jacobian maps: 91.06% on ADNI and 94.3% on OASIS-3. If it holds up, this gives neuroimaging pipelines a much cheaper preprocessing step for large-scale dementia screening.","feed_headline":"Brain scans: gradient-angle maps match Jacobian maps at 6x lower cost","feed_subtitle":"New preprocessing method reaches comparable Alzheimer's classification accuracy in 93 ms instead of 584 ms per image.","key_machinery":"The central object is the Sobel kernel angle difference (SKAD), computed in Eq. 7 as $\\left|\\tan^{-1}\\left(\\frac{\\sqrt{\\nabla M_x^2 + \\nabla M_y^2}}{\\nabla M_z}\\right) - \\tan^{-1}\\left(\\frac{\\sqrt{\\nabla F_x^2 + \\nabla F_y^2}}{\\nabla F_z}\\right)\\right|$, where $\\nabla M$ and $\\nabla F$ are the 3D Sobel gradient vectors of the registered moving image and the fixed template. The 3x3x3 Sobel kernels replace explicit finite-difference derivatives with convolution, and the angle between gradient vectors is used as a scalar summary of local structural orientation disparity at each voxel. The machinery's job is to produce a single map per brain that highlights regions where the patient's image and the template disagree in gradient orientation, which the paper argues marks atrophy.","core_discovery":"SKAD is defined by Eq. 7 as the absolute difference between the 3D gradient-direction angles of the warped patient image and the fixed MNI template, where gradients are obtained by convolving the image volume with 3x3x3 Sobel kernels in the x, y, and z directions. The paper's central claim is that this angle difference captures regional volume variations and signifies abnormalities or atrophy, making it a viable substitute for the Jacobian determinant map, which is the standard measure of local volume change during nonlinear registration. On the ADNI and OASIS-3 datasets, a VGG-16 model fed SKAD maps achieves 91.06% and 94.3% accuracy respectively, compared with 94.0% and 89.4% for Jacobian maps in the paper's own ablation, while SKAD costs 93 ms and 1.66 billion FLOPs versus 584 ms and 146 billion FLOPs for the Jacobian map. The authors conclude that SKAD is an efficient and competitive approach for dementia diagnosis.","pith_inferences":["A direct voxelwise comparison of SKAD against log-Jacobian determinants would clarify whether SKAD truly measures volume change or instead captures edge and boundary shifts; the paper does not report such a comparison.","Because SKAD only compares gradient orientations, it is insensitive to the magnitude of tissue loss; a region that shrinks while preserving its edge orientation could produce little or no SKAD signal, which would limit sensitivity in early atrophy.","SKAD's speed makes it a candidate preprocessing step for large self-supervised models that need many images; the paper notes this possibility but does not test it.","The angle difference could be computed earlier, directly from the deformation field, which might isolate the orientation-change component; that variant is not explored in the paper."],"forward_implications":["A VGG-16 model trained on SKAD inputs reaches 91.06% accuracy on ADNI and 94.3% on OASIS-3, on par with Jacobian inputs in the paper's experiments.","Per-image preprocessing drops from 584 ms to 93 ms and from about 146 billion to 1.66 billion floating-point operations, making large-scale and repeated imaging analysis much cheaper.","Because SKAD has lower memory use (1199.7 MiB vs 1267.3 MiB for Jacobian maps), it can be applied to larger batches or on more modest hardware.","The method is not tied to Alzheimer's specifically and could extend to other neurodegenerative diseases where atrophy patterns matter, such as Parkinson's, if relevant imaging data are available."],"supporting_citations":[{"why":"introduces Jacobian determinants for VBM modulation, the volume-change measure SKAD is designed to replace.","marker":"[6]"},{"why":"supplies the nonlinear registration and bias-field correction tools that produce the warped images used by both methods.","marker":"[7]"},{"why":"provides the OASIS-3 structural MRI dataset used for evaluation.","marker":"[20]"},{"why":"provides the ADNI-I dataset used for evaluation.","marker":"[28]"},{"why":"presents a Jacobian-domain CNN whose accuracy serves as a state-of-the-art comparison.","marker":"[1]"},{"why":"presents a Jacobian-map multimodal model used as a comparison baseline.","marker":"[22]"},{"why":"presents a Jacobian-guided CNN explainability method, also used as a comparison baseline.","marker":"[23]"}],"fun_headline_variants":["SKAD: 6.3x faster Alzheimer's detection with similar accuracy","New angle-based method speeds brain imaging 6.3x for dementia","Gradient angle maps cut Alzheimer's scan analysis cost 6x","Faster Alzheimer's detection: gradient angles rival Jacobian maps","SKAD: Alzheimer's detection at 6x speed, comparable accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the difference in angle between the Sobel gradient orientations of a warped patient image and the fixed healthy template is a valid indicator of local brain-volume loss, even though the paper does not derive or empirically validate the link between this angle difference and the Jacobian determinant, which is the actual measure of volume change.","fun_headline_variants_meta":{"raw":{"variants":["SKAD: 6.3x faster Alzheimer's detection with similar accuracy","New angle-based method speeds brain imaging 6.3x for dementia","Gradient angle maps cut Alzheimer's scan analysis cost 6x","Faster Alzheimer's detection: gradient angles rival Jacobian maps","SKAD: Alzheimer's detection at 6x speed, comparable accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000352,"raw_usage":{"total_tokens":1909,"prompt_tokens":930,"completion_tokens":979,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":885}},"tokens_in":546,"tokens_out":979,"duration_ms":9066,"temperature":1.0,"reasoning_tokens":885,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:20:37.808602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute both the SKAD map and the Jacobian determinant map from the same set of registered MRI scans and measure the voxelwise correlation between SKAD values and log-Jacobian determinants. If the correlation is near zero, or if SKAD's classification performance comes primarily from template registration artifacts rather than from true volume change, then the claim that SKAD captures regional volume variations is falsified.","supporting_citations":[{"cited_title":"V oxel-based morphome- try—the methods","cited_arxiv_id":null,"evidence_quote":"introduces Jacobian determinants for VBM modulation, the volume-change measure SKAD is designed to replace."},{"cited_title":"Advanced normalization tools (ants)","cited_arxiv_id":null,"evidence_quote":"supplies the nonlinear registration and bias-field correction tools that produce the warped images used by both methods."},{"cited_title":"Oasis-3: longitudinal neuroimaging, clin- ical, and cognitive dataset for normal aging and alzheimer disease","cited_arxiv_id":null,"evidence_quote":"provides the OASIS-3 structural MRI dataset used for evaluation."},{"cited_title":"The alzheimer’s disease neuroimaging initiative: progress report and future plans","cited_arxiv_id":null,"evidence_quote":"provides the ADNI-I dataset used for evaluation."},{"cited_title":"Transformed domain convolutional neural network for alzheimer’s disease diagnosis using structural mri","cited_arxiv_id":null,"evidence_quote":"presents a Jacobian-domain CNN whose accuracy serves as a state-of-the-art comparison."},{"cited_title":"Diagnosing Alzheimer’s disease using early-late multimodal data fusion with Jacobian maps","cited_arxiv_id":null,"evidence_quote":"presents a Jacobian-map multimodal model used as a comparison baseline."},{"cited_title":"Unmasking dementia detec- tion by masking input gradients: A JSM approach to model interpretability and precision","cited_arxiv_id":null,"evidence_quote":"presents a Jacobian-guided CNN explainability method, also used as a comparison baseline."}],"review_version":1}