Distance Transform Guided Mixup for Alzheimer's Detection

Erchan Aptoula; Huseyin Ozkan; Zobia Batool

REVIEW 5 major objections 5 minor 12 references

Distance Transform Guided Mixup for Alzheimer's Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that distance-transform-guided mixup, which swaps anatomy-defined layers between MRI scans, improves single-domain generalization for Alzheimer's classification, achieving the best accuracy and F1 on ADNI and strong…

desk verdict A structurally motivated mixup variant for 3D MRI that is worth a look, but the reported F1 gains are confounded with an unshared class-weighted loss and the abstract overstates consistency across ADNI and AIBL. read the letter →

arxiv 2505.22434 v1 pith:L5KBD7KP submitted 2025-05-28 cs.CV

Zobia Batool , Huseyin Ozkan , Erchan Aptoula This is my paper

classification cs.CV

keywords Alzheimer'sdiseaseclassificationdomaingeneralizationdistancetransformdataaugmentationmixup3DMRIsingle-domainmedicalimageanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a data augmentation method for Alzheimer's disease classification from 3D MRI scans, aimed at the single-domain generalization setting where a model is trained on one dataset and tested on others. Instead of blending entire images as standard mixup does, the method computes a distance transform of each scan, splits the brain into four spatial layers by distance to the nearest boundary, and swaps these layers between pairs of scans to create mixed training images. The label of a mixed image is the pixel-count-weighted combination of the two source labels, trained with a weighted soft cross-entropy loss. On the ADNI external test set the method reaches 48.37% accuracy and an F1 of 0.460, the best among compared baselines, and on AIBL it reaches 52.25% accuracy with the highest F1 (0.430). The authors conclude that structure-aware, region-based mixing improves generalization while preserving brain anatomy.

What carries the argument

Distance transform guided mixup: for each MRI, $D(p) = \min_{q \in B} \|p - q\|$ gives every voxel its Euclidean distance to the nearest background point; thresholds $t_1$ and $t_2$ split the brain into four mutually exclusive binary masks $R_1, R_2, R_3, R_4$. Mixed images are formed by taking $R_1$ and $R_3$ from the first scan and $R_2$ and $R_4$ from the second (Eq. 6), and labels are mixed by the voxel-count fractions from the two sources (Eqs. 7-10). This replaces pixel-wise linear interpolation with spatial swapping of whole anatomical layers.

What would settle it

Compare the proposed distance-transform mixing against a control that splits each brain into four regions using random axis-aligned quadrants or concentric shells instead of distance thresholds, keeping everything else identical. If the control matches the proposed method on ADNI and AIBL accuracy, the distance-transform guidance itself is not responsible for the improvement; if the control is worse, the anatomical-layering hypothesis is supported. Additionally, an expert or a validated separate AD classifier can rate mixed images: if mixed images frequently contradict their pixel-weighted labels (e.g., an image that is mostly healthy tissue but labeled with a large AD share), the label-mixing formula is corrupting supervision.

Watch

Extended reading notes

Core claim

The paper's central claim is that replacing the random pixel-blending of standard mixup with a structure-aware blend of distance-transform layers preserves disease-relevant brain anatomy while still generating diverse training samples, and that this improves a classifier's ability to generalize from one training dataset to unseen test datasets. The method computes for each scan a distance transform, thresholds it into four nested regions, and constructs two mixed scans per pair by assigning alternating regions to each source; the soft training label is the pixel-count-weighted combination of the source labels. In experiments, this approach reaches 48.37% accuracy and a 0.460 F1 on ADNI—the best among the compared methods—and 52.25% accuracy with a 0.430 F1 on AIBL, where it also has the highest F1. The authors frame this as evidence that region-aware, structure-preserving augmentation is a viable single-domain generalization strategy for Alzheimer's MRI classification.

Load-bearing premise

The load-bearing premise is that the distance-transform layers are semantically meaningful partitions of brain anatomy, so that swapping layers between subjects preserves disease-relevant cues and keeps the pixel-weighted label a valid training signal.

Editorial extensions

If this is right

The augmentation adds only an offline distance-transform computation and fixed masking operations, so it can be inserted into any 3D classification pipeline without extra trainable parameters.
On both external test sets, the method outperforms the plain pre-trained U-Net baseline by more than 10 accuracy points, indicating that the mixing strategy contributes beyond the backbone initialization.
Because the method requires only a single training domain, it directly addresses the single-domain generalization setting without collecting additional source datasets.
The authors state the same augmentation strategy can be applied to other neurodegenerative diseases that experience similar cross-dataset domain shift.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A control experiment that replaces distance-transform layers with geometrically defined shells or random quadrants would isolate whether the distance-transform semantics or simply the spatial locality of mixing drives the gains.
The label-mixing rule weights every voxel equally, so small-volume disease markers such as hippocampal atrophy could be diluted; a label rule weighted by regional disease relevance is a testable alternative.
The reported ADNI accuracy is below 50%, and no confidence intervals are given; repeated runs with different seeds and splits would establish whether the ranking over baselines is stable.
Since all scans are registered to MNI152 before mixing, a sensitivity analysis with perturbed registration would reveal how much of the benefit depends on precise anatomical alignment.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the distance-transform-guided mixing is a genuinely new twist on mixup for 3D MRI, and the paper does the right thing by training on NACC and testing on ADNI/AIBL. But the headline claim is not supported by the numbers as reported. On ADNI the accuracy edge over plain Mixup is 0.08 percentage points; on AIBL the method is 13 points worse than Mixup in accuracy, and the abstract's “improvement across both datasets” is only true if you choose F1. The bigger problem is the evaluation protocol: the proposed pipeline adds inverse-frequency class weights to the soft cross-entropy (Eq. 11), and Section III-B never says Mixup, RSC, CCSDG, or the U-Net baseline use the same weighted loss. ADNI and especially AIBL are heavily imbalanced (AIBL: 465 NC vs 68 AD), so reweighting alone can lift macro-F1 and sensitivity while hurting accuracy. That is exactly the AIBL pattern. Without controlling for the loss, you cannot attribute the F1 gains to the distance-transform mixing.

What is genuinely new: the idea of thresholding distance transforms to produce four non-overlapping region masks and mixing those regions rather than blending voxels globally. That is structure-preserving in a way plain mixup is not, and it is not in the cited literature. The external evaluation design is also right: train on NACC, test on 80% of ADNI/AIBL. The method is simple and reproducible from the description, apart from a couple of details.

The remaining soft spots are in proportion. No error bars or repeated seeds, so the ADNI accuracy difference (48.37 vs 48.29) is one subject of noise. The label mixing weights voxel counts from each region equally (Eq. 10); no evidence that voxel count tracks disease-relevant content, which is an assumption worth testing. The 10% preservation ratio is presented without sensitivity analysis. And on AIBL the accuracy gap to Mixup is large enough to make the “outperforms baselines” sentence in the conclusion wrong for that metric.

Who this is for: someone building augmentation for volumetric medical images; they will want to re-run the comparison with a matched loss and add seeds. The paper deserves a serious referee, but it needs revision before the generalization claim can be believed. If I were handling it, I'd send it out with a request for a controlled baseline comparison and error bars.

Referee Report

5 major / 5 minor

Summary. The paper proposes a single-domain generalization method for Alzheimer's disease classification from 3D MRI. A distance transform of each scan is thresholded into four spatial regions, and mixed samples are formed by combining regions from two scans. Soft labels are computed as the pixel-count-weighted sum of the source labels, and training uses a 3D U-Net with weighted soft cross-entropy loss on the NACC dataset. The method is evaluated on ADNI and AIBL and compared with MixUp, RSC, CCSDG, and a 3D U-Net baseline. The authors claim improved generalization across both external datasets.

Significance. If the claims were fully supported, the paper would provide a simple, structure-aware augmentation for medical imaging that requires no extra source domains and could transfer to other neurodegenerative diseases. The use of public datasets (NACC, ADNI, AIBL) and comparison with several baselines is a strength. However, the current experimental evidence does not establish the central claim: the ADNI accuracy margin over MixUp is negligible, the AIBL accuracy is substantially below MixUp, and the comparison is confounded by the class-weighted loss. The method's core assumption that distance-transform regions preserve disease-relevant structure is also unvalidated. These issues are fixable, but they require new experiments and a revised presentation.

major comments (5)

[Section III-B/III-C, Eq. (11), Tables II and III] The comparison is confounded by the class-weighted loss. The paper does not state whether MixUp, RSC, CCSDG, or the 3D U-Net baseline use the same weighted soft cross-entropy defined in Eq. (11). Because AIBL is heavily imbalanced (465 NC, 101 MCI, 68 AD), inverse-frequency reweighting alone can raise macro-F1 and sensitivity while lowering accuracy. This is exactly the observed AIBL pattern: the proposed method has higher F1 (0.430 vs 0.382) but much lower accuracy (52.25% vs 65.42%) than MixUp. Without controlling the loss function across all methods, the reported F1/sensitivity gains cannot be attributed to the distance-transform mixing rather than to reweighting. The authors should rerun all baselines with the identical weighted loss or report an unweighted variant of their own method.
[Table II and Section III-C] The ADNI accuracy advantage over MixUp is 48.37% vs 48.29%, a difference of 0.08 percentage points, which on the approximately 1,258-case evaluation set corresponds to about one subject. No repeated-seed statistics, standard deviations, confidence intervals, or significance tests are reported. The claim that the method 'outperformed all baselines' in accuracy is not supported by a difference of this size. The authors should report mean and standard deviation over multiple seeds and, where appropriate, a paired significance test.
[Abstract, Section III-C, Table III] The abstract claims 'generalization performance improvement across both ADNI and AIBL datasets,' but Table III shows that MixUp achieves 65.42% accuracy on AIBL while the proposed method achieves only 52.25%. The text in Section III-C acknowledges only 'competitive results' on AIBL. The stated claim is therefore contradicted by the paper's own table. Either the claim must be revised to specify which metric improves on AIBL, or the authors must explain why a large accuracy drop is acceptable given the method's objective. If macro-F1 is the primary metric, this should be stated explicitly and justified.
[Section II-B, Eqs. (2)-(5)] The region construction is not precisely specified and likely not a coherent partition. R1 is defined using Da only, R2 using Db only, R3 using Da only, and R4 using Db only, with recursive exclusions that mix distance transforms from the two images. It is unclear how thresholds t1 and t2 are chosen for a given pair, how the 'at least 10% of the brain structure is preserved' condition is enforced, and whether the masks are defined in a common coordinate space after registration. Without a precise algorithm, the method cannot be reproduced, and the claimed preservation of brain structure cannot be evaluated. The authors should give an exact, step-by-step definition of the four masks and the threshold selection procedure, ideally with a diagram showing the actual masks on an MRI.
[Section II-C, Eq. (10)] The label-mixing rule assumes that every voxel contributes equally to the diagnosis and that the four distance-transform regions are semantically meaningful partitions of brain anatomy. This is not validated. If a region containing disease-relevant atrophy is swapped between a normal and an AD scan, Eq. (10) may assign a label that does not correspond to the actual pathological content of the mixed image. No analysis is provided to show that mixed images retain the cues needed for classification, and no sensitivity analysis examines the effect of the 10% preservation ratio. A control experiment or visual/quantitative inspection of mixed samples would strengthen the claim that the augmentation preserves label validity.

minor comments (5)

[References] Reference [4] contains the placeholder 'Insert Page Numbers' and reference [5] is listed only as 'arXiv preprint, 2017' without the arXiv identifier; both should be completed.
[Eq. (11)] The denominator '1P wi' appears to be a typo for 'sum_i w_i'; please correct the formula and also state explicitly how the inverse-frequency weights are computed.
[Eq. (1) and throughout] The text uses 'pixel' for what are three-dimensional voxels; please use 'voxel' consistently for the 3D MRI data.
[Table I and Section III-A] The paper says the model is evaluated on '80% of the target datasets,' but Table I appears to report full-dataset counts; please clarify whether the demographic counts correspond to the full datasets or the 80% evaluation subsets.
[Section III-B] No hyperparameter sensitivity analysis is provided for the 10% preservation ratio or the threshold choice; since these parameters control the augmentation, a sensitivity study or at least a brief discussion would help the reader understand their impact.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the distance-transform mixup is an empirical augmentation method with no fitted parameter or self-citation chain driving its claimed result.

full rationale

The paper's derivation chain is fully explicit and non-circular. The distance transform D(p) of Eq. 1 is a standard geometric computation from the input image; thresholds t1 and t2 are set per image only to preserve at least 10% of the brain structure per region, which is a stated design choice rather than a parameter fitted to target performance. Region masks (Eqs. 2-5) and the mixed image (Eq. 6) are deterministic functions of the two input scans. The soft label in Eq. 10 is computed from the pixel-count proportions in Eqs. 7-9, so the label is by construction consistent with the mixing rule; this is not a circular prediction because no claim is made that the label formula is derived from, or fitted to, the outcome labels of ADNI or AIBL. The class weights in Eq. 11 are inverse-frequency weights, again a standard design choice, and the paper reports held-out performance on ADNI and AIBL from a model trained on NACC. The backbone is initialized with chest-CT pretrained weights from Models Genesis [10], an external source, and the comparison baselines are external methods. No load-bearing argument reduces to a self-citation, and no result is renamed or imported from the authors' prior work. The skeptical concern that the class-weighted loss is not stated to be shared by baselines is an experimental confounding risk, not circularity; likewise the small ADNI accuracy margin is a statistical robustness issue, not a definitional equivalence. Therefore no significant circularity is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method rests on the unvalidated assumption that distance-to-boundary layers are semantically meaningful for Alzheimer's pathology and that mixing them across subjects preserves label information. The only hand-chosen free parameter is the 10% region preservation ratio; the class weights and loss are standard formulas.

free parameters (1)

minimum region preservation ratio (10%) = 10% (hand-chosen)
The paper requires each of the four distance-transform regions to preserve at least 10% of the brain structure (Section II-B), but does not explain how this constraint maps to thresholds t1 and t2, nor does it justify the specific percentage. This is a hand-chosen design hyperparameter.

assumptions (3)

domain assumption Euclidean distance transform from background voxels provides a meaningful one-dimensional ordering of brain anatomy.
The method (Eq. 1) assumes that distance to the nearest background captures structure in a way that makes layer-based mixing useful. This is not validated in the paper.
domain assumption Mixing distance-transform regions from two subjects yields training images whose soft labels (Eq. 10) remain valid for the classification task.
The label mixing is based on voxel counts (Eqs. 7-9), implicitly assuming that each voxel's contribution to the label is uniform. No evidence is given that the mixed images are semantically plausible for AD detection.
ad hoc to paper The 10% preservation rule produces regions of adequate size without hurting diversity.
Stated in Section II-B to 'avoid creating excessively small regions', but the threshold is not derived or tested against alternatives.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distance Transform Guided Mixup for Alzheimer's Detection." pith.science (2026). https://pith.science/paper/L5KBD7KP

@misc{pith2026250522434,
  author       = {Pith},
  title        = {Pith review of: Distance Transform Guided Mixup for Alzheimer's Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5KBD7KP}},
  note         = {Machine review of arXiv:2505.22434}
}

read the original abstract

Alzheimer's detection efforts aim to develop accurate models for early disease diagnosis. Significant advances have been achieved with convolutional neural networks and vision transformer based approaches. However, medical datasets suffer heavily from class imbalance, variations in imaging protocols, and limited dataset diversity, which hinder model generalization. To overcome these challenges, this study focuses on single-domain generalization by extending the well-known mixup method. The key idea is to compute the distance transform of MRI scans, separate them spatially into multiple layers and then combine layers stemming from distinct samples to produce augmented images. The proposed approach generates diverse data while preserving the brain's structure. Experimental results show generalization performance improvement across both ADNI and AIBL datasets.

Figures

Figures reproduced from arXiv: 2505.22434 by the authors.

**Figure 1.** Overview of the proposed Alzheimer’s disease classification pipeline. 3D MRI scans are preprocessed before applying region mixing augmentation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Overview of the mixing strategy. Given two input MRI scans ( [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Recent advances in alzheimer’s disease: Mechanisms, clinical trials and new drug development strategies,

J. Zhang, Y . Zhang, J. Wang, J. Xia, J. Zhang, and L. Chen, “Recent advances in alzheimer’s disease: Mechanisms, clinical trials and new drug development strategies,” Signal Transduction and Targeted Ther- apy, vol. 9, no. 1, p. 211, Aug. 2024

work page 2024
[2]

Disease-driven domain generalization for neuroimaging- based assessment of alzheimer’s disease,

D. Lteif, S. Sreerama, S. A. Bargal, B. A. Plummer, R. Au, and V . B. Kolachalama, “Disease-driven domain generalization for neuroimaging- based assessment of alzheimer’s disease,” Human Brain Mapping , vol. 45, no. 8, p. e26707, Jun. 2024

work page 2024
[3]

Medtransformer: Accurate ad diagnosis for 3d mri images through 2d vision transformers,

Y . Wang, K. Chen, Y . Zhang, and H. Wang, “Medtransformer: Accurate ad diagnosis for 3d mri images through 2d vision transformers,” arXiv preprint, Jan. 2024

work page 2024
[4]

Learning with domain- knowledge for generalizable prediction of alzheimer’s disease from multi-site structural mri,

Y . Zhou, Y . Li, F. Zhou, Y . Liu, and L. Tu, “Learning with domain- knowledge for generalizable prediction of alzheimer’s disease from multi-site structural mri,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2023 , ser. Lecture Notes in Computer Science, H. G. et al., Ed. Cham, Switzerland: Springer, 2023, vol. 14224, p. Insert...

work page 2023
[5]

Mixup: Beyond empirical risk minimization,

H. Zhang, “Mixup: Beyond empirical risk minimization,” arXiv preprint, 2017

work page 2017
[6]

Domain generalization with mixstyle,

K. Zhou, Y . Yang, Y . Qiao, and T. Xiang, “Domain generalization with mixstyle,” arXiv preprint, Apr. 2021

work page 2021
[7]

The national alzheimer’s coordinating center (nacc) database: an alzheimer disease database,

D. L. Beekly, E. M. Ramos, G. van Belle, W. Deitrich, A. D. Clark, M. E. Jacka, and W. A. Kukull, “The national alzheimer’s coordinating center (nacc) database: an alzheimer disease database,” Alzheimer Disease & Associated Disorders, vol. 18, no. 4, pp. 270–277, 2004

work page 2004
[8]

Alzheimer’s disease neuroimaging initiative (adni) clinical characterization,

R. C. Petersen, P. S. Aisen, L. A. Beckett, M. C. Donohue, A. C. Gamst, D. J. Harvey, C. R. J. Jr, W. J. Jagust, L. M. Shaw, A. W. Toga, and J. Q. Trojanowski, “Alzheimer’s disease neuroimaging initiative (adni) clinical characterization,” Neurology, vol. 74, no. 3, pp. 201–209, 2010

work page 2010

Show all 12 references

[9]

K. A. Ellis, A. I. Bush, D. Darby, D. D. Fazio, J. Foster, P. Hudson, N. T. Lautenschlager, N. Lenzo, R. N. Martins, R. Maruff, P. Masters, and the AIBL Research Group, “The australian imaging, biomarkers and lifestyle (aibl) study of aging: methodology and baseline characteri...

2009
[10]

Models genesis: Generic autodidactic models for 3d medical image analysis,

Z. Zhou, V . Sodha, M. M. R. Siddiquee, R. Feng, N. Tajbakhsh, M. B. Gotway, and J. Liang, “Models genesis: Generic autodidactic models for 3d medical image analysis,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2019 . Cham: Springer International Pu...

2019
[11]

Self-challenging im- proves cross-domain generalization,

Z. Huang, H. Wang, E. P. Xing, and D. Huang, “Self-challenging im- proves cross-domain generalization,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, Aug. 23–28, 2020, Proceed- ings, Part II 16 . Springer International Publishing, 2020, p. 124–140

2020
[12]

Devil is in channels: Contrastive single domain generalization for medical image segmentation,

S. Hu, Z. Liao, and Y . Xia, “Devil is in channels: Contrastive single domain generalization for medical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Inter- vention. Springer, 2023, pp. 14–23

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Recent advances in alzheimer’s disease: Mechanisms, clinical trials and new drug development strategies,

J. Zhang, Y . Zhang, J. Wang, J. Xia, J. Zhang, and L. Chen, “Recent advances in alzheimer’s disease: Mechanisms, clinical trials and new drug development strategies,” Signal Transduction and Targeted Ther- apy, vol. 9, no. 1, p. 211, Aug. 2024

work page 2024

[2] [2]

Disease-driven domain generalization for neuroimaging- based assessment of alzheimer’s disease,

D. Lteif, S. Sreerama, S. A. Bargal, B. A. Plummer, R. Au, and V . B. Kolachalama, “Disease-driven domain generalization for neuroimaging- based assessment of alzheimer’s disease,” Human Brain Mapping , vol. 45, no. 8, p. e26707, Jun. 2024

work page 2024

[3] [3]

Medtransformer: Accurate ad diagnosis for 3d mri images through 2d vision transformers,

Y . Wang, K. Chen, Y . Zhang, and H. Wang, “Medtransformer: Accurate ad diagnosis for 3d mri images through 2d vision transformers,” arXiv preprint, Jan. 2024

work page 2024

[4] [4]

Learning with domain- knowledge for generalizable prediction of alzheimer’s disease from multi-site structural mri,

Y . Zhou, Y . Li, F. Zhou, Y . Liu, and L. Tu, “Learning with domain- knowledge for generalizable prediction of alzheimer’s disease from multi-site structural mri,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2023 , ser. Lecture Notes in Computer Science, H. G. et al., Ed. Cham, Switzerland: Springer, 2023, vol. 14224, p. Insert...

work page 2023

[5] [5]

Mixup: Beyond empirical risk minimization,

H. Zhang, “Mixup: Beyond empirical risk minimization,” arXiv preprint, 2017

work page 2017

[6] [6]

Domain generalization with mixstyle,

K. Zhou, Y . Yang, Y . Qiao, and T. Xiang, “Domain generalization with mixstyle,” arXiv preprint, Apr. 2021

work page 2021

[7] [7]

The national alzheimer’s coordinating center (nacc) database: an alzheimer disease database,

D. L. Beekly, E. M. Ramos, G. van Belle, W. Deitrich, A. D. Clark, M. E. Jacka, and W. A. Kukull, “The national alzheimer’s coordinating center (nacc) database: an alzheimer disease database,” Alzheimer Disease & Associated Disorders, vol. 18, no. 4, pp. 270–277, 2004

work page 2004

[8] [8]

Alzheimer’s disease neuroimaging initiative (adni) clinical characterization,

R. C. Petersen, P. S. Aisen, L. A. Beckett, M. C. Donohue, A. C. Gamst, D. J. Harvey, C. R. J. Jr, W. J. Jagust, L. M. Shaw, A. W. Toga, and J. Q. Trojanowski, “Alzheimer’s disease neuroimaging initiative (adni) clinical characterization,” Neurology, vol. 74, no. 3, pp. 201–209, 2010

work page 2010