Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Attention-based FLAIR segmentation plus simple shape features can separate vascular from demyelinating white-matter lesions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 11:55 UTC pith:ZFLZHJT3

load-bearing objection Solid multi-site ablation of known attention modules for WMH segmentation plus shape-feature etiology classification; the classification claim is the softest part because labels track datasets. the 3 major comments →

arxiv 2607.08171 v1 pith:ZFLZHJT3 submitted 2026-07-09 cs.CV

Attention-Based Segmentation of WMHs and Differentiation of Vascular vs. Demyelinating Lesions

Aina Tur-Serrano , Gabriel Moy\`a-Alcover , Francisco J. Perales L\'opez This is my paper
classification cs.CV
keywords white matter hyperintensitiesattention U-NetFLAIR MRIlesion segmentationmorphological featuresvascular lesionsdemyelinating lesionsBAM CBAM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

White-matter hyperintensities look nearly identical on FLAIR MRI whether they arise from vascular disease or demyelinating conditions, so radiologists often struggle to tell the causes apart. This paper argues that attention modules inside U-Net architectures improve lesion segmentation, after which a short list of geometric measurements taken from the resulting masks is enough for ordinary classifiers to label each lesion by origin. Training on ordinary 2-D slices beats both patch-based and limited 2.5-D inputs, and classification accuracy obtained from the automatic masks stays close to accuracy obtained from expert masks. Five public datasets that differ in scanner, resolution and patient mix are used to show the pipeline still works under realistic heterogeneity. If the pattern holds, clinics could obtain faster, more consistent differential support without multi-modal imaging or large end-to-end networks for the final decision.

Core claim

The authors establish that attention-augmented segmentation (Attention U-Net plus BAM and CBAM) followed by extraction of morphological descriptors from connected lesion components yields a two-step pipeline that discriminates vascular from demyelinating white-matter hyperintensities on FLAIR. Classification performance on the predicted masks remains close to performance on ground-truth masks across five heterogeneous public datasets, even though sample sizes are modest.

What carries the argument

The two-step pipeline of attention-enhanced U-Net segmentation of FLAIR lesions followed by extraction of 2-D morphological features (area, eccentricity, roundness, solidity and related shape measures) that train shallow classifiers.

Load-bearing premise

A short list of two-dimensional shape measurements taken from lesion masks remains informative enough to separate the two causes even when scanners, resolutions and patient groups differ.

What would settle it

On a fresh multi-center FLAIR cohort with confirmed vascular and demyelinating diagnoses, extract the same morphological features from expert masks and retrain the classifiers; if accuracy falls to chance or the vascular-demyelinating gap vanishes, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • 2-D slice training with attention modules can be preferred over patch or 2.5-D strategies when data volumes are limited.
  • Shape and location features alone carry usable signal for etiology once lesions are segmented.
  • Small segmentation errors do not destroy downstream classification accuracy.
  • Attention placed throughout the network improves detection of lesion-free control scans.
  • The same lightweight pipeline can be re-validated on larger clinical cohorts for routine differential support.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the morphological descriptors stay stable under domain shift, any existing WMH segmenter could add the same feature-classification step as a lightweight, auditable post-process.
  • Reconstructing 3-D shape descriptors by stacking the 2-D masks, as the authors themselves flag for future work, would likely sharpen discrimination of elongated periventricular lesions.
  • The approach may transfer to other look-alike lesion pairs where radiologists already rely on shape and location cues.
  • Because the final classifiers are shallow and feature-based, clinical bias can be inspected more easily than with end-to-end deep models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a two-stage pipeline for white-matter hyperintensity (WMH) analysis on FLAIR MRI: (1) attention-enhanced U-Net variants (Attention U-Net, BAM, CBAM placed in encoder/decoder/full) trained under 2D-slice, patch-based and 2.5D regimes, and (2) extraction of 2-D morphological descriptors (area, eccentricity, roundness, solidity, centroid position, etc.) from connected components of the resulting masks, followed by shallow classifiers (SVM, logistic regression, random forest) that assign each lesion a vascular versus demyelinating label. Five public multi-scanner datasets (vascular, MS and healthy controls) are used under five-fold volume-wise cross-validation. The strongest empirical claim is that 2D Attention U-Net + full BAM/CBAM yields the highest overall Dice (~0.72) while plain Attention U-Net is more lesion-sensitive, that classification accuracy on predicted masks remains close to accuracy on ground-truth masks (Table 8, ~0.60–0.75), and that the combination therefore constitutes a “promising direction” for etiology discrimination.

Significance. If the morphological features truly capture etiology rather than residual dataset/scanner cues, the work would supply an interpretable, lightweight alternative to end-to-end deep classifiers for a clinically relevant differential diagnosis that is currently difficult on FLAIR alone. The systematic comparison of attention modules and input regimes on heterogeneous public data, together with the explicit GT-versus-predicted classification table, is a useful empirical contribution even if the absolute numbers remain modest. The authors correctly flag the need for larger clinical cohorts.

major comments (3)
  1. §2.1–2.2, Table 1 and §4.2 / Table 8: every lesion inherits its binary etiology label from the source dataset (vascular cohorts vs. MS cohorts). The five datasets differ systematically in age, field strength, resolution and preprocessing. Because 5-fold CV only keeps volumes together and never isolates a whole dataset or scanner, the reported classification accuracies (0.60–0.75) could be driven by residual domain cues that survive the simple morphological pipeline rather than by true vascular-versus-demyelinating shape differences. A leave-one-dataset-out or scanner-matched ablation is required to substantiate the claim that the descriptors are etiology-discriminative across heterogeneous scanners and populations.
  2. Tables 3–4 and 7: Dice on demyelinating lesions is consistently ~0.10–0.15 lower than on vascular lesions for every architecture and input regime. Because classification is performed only on slices that contain at least one lesion and uses the same predicted masks, the poorer segmentation of the harder class systematically biases the feature distributions fed to the classifiers. The manuscript does not quantify how much of the demyelinating-class performance drop (Table 8) is attributable to this segmentation gap versus intrinsic morphological overlap.
  3. §3.5 and Table 2: the per-fold validation sets are small (12–34 volumes). With only five folds and high inter-scanner heterogeneity, the reported standard deviations (especially for Attention U-Net variants) are large enough that many of the claimed ranking differences among attention placements fall inside the error bars. Statistical tests or confidence intervals on the Dice differences would be needed before asserting that full BAM+CBAM is superior.
minor comments (5)
  1. Table 1 header “Number of scanners” actually lists number of volumes; the true scanner counts are never stated.
  2. §2.2: Otsu thresholding + hole-filling is applied after slice extraction; the effect of this step on small or low-contrast lesions is not quantified.
  3. Figure 1 caption and surrounding text do not indicate which fold or which attention configuration is shown, making visual comparison difficult.
  4. §3.3: hyper-parameters for the three classifiers were obtained by randomized search, yet the search ranges and the final selected values for every fold are not reported.
  5. Several references (e.g., [22]–[25], [27]–[28]) appear only loosely related to the technical claims and could be pruned or better contextualized.

Circularity Check

0 steps flagged

No circularity: purely empirical pipeline with independent ground-truth evaluation; morphological features and classifiers do not reduce to inputs by construction.

full rationale

The paper is an empirical medical-imaging study. Segmentation models (U-Net / Attention U-Net + BAM/CBAM variants) are trained with Dice loss against independent manual lesion masks drawn from five public datasets; performance is reported via Dice, Jaccard, precision and recall on held-out volumes (Tables 3–7). Morphological descriptors (area, perimeter, solidity, roundness, eccentricity, etc.) are extracted from connected components of those masks by a fixed geometric library and are never defined in terms of the etiology label. Binary labels (vascular vs. demyelinating) are inherited from the source dataset and are used only for supervised training of shallow classifiers (SVM, LR, RF); classification metrics are computed on both ground-truth and predicted masks (Table 8). No equation, fitted constant or uniqueness theorem is claimed to “derive” a prediction; the only self-reference is the authors’ open-source feature library, which is a computational tool, not a load-bearing theoretical premise. Consequently the derivation chain contains no self-definitional, fitted-input-as-prediction, or self-citation-load-bearing steps. Domain-shift confounds may affect correctness, but they do not constitute circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard deep-learning practice plus domain assumptions about FLAIR sensitivity and the clinical meaning of lesion shape. Free parameters are ordinary training hyper-parameters chosen by hand or randomized search; no new physical constants or invented entities are introduced.

free parameters (5)
  • Adam learning rate = 1e-5
    Fixed at 1e-5 for all experiments; not derived from first principles.
  • training epochs = 15-30
    20 for 2-D, 30 for 2.5-D, 15-20 for patches; chosen by observed convergence.
  • patch size and overlap = 64x64 / 16
    64x64 with 16-pixel overlap selected by the authors.
  • SVM C and gamma = C=10, gamma=0.1
    Obtained by randomized search on the training folds.
  • Random Forest n_estimators and max_depth = 200 / 20
    Set to 200 trees, depth 20 after randomized search.
axioms (4)
  • domain assumption FLAIR is the most sensitive MRI sequence for detecting WMHs
    Stated in Introduction and Section 2.1; used to justify single-modality input.
  • domain assumption Morphological descriptors (eccentricity, roundness, solidity, etc.) capture clinically meaningful differences between vascular and demyelinating lesions
    Section 2.4 links features to radiological criteria; load-bearing for the classification claim.
  • domain assumption Public dataset lesion annotations constitute reliable ground truth for both segmentation and etiology labels
    Assumed throughout training and evaluation; no inter-rater reliability numbers supplied.
  • standard math Standard U-Net / Attention U-Net training with Dice loss and Adam is a valid optimization procedure
    Background deep-learning practice used without re-derivation.

pith-pipeline@v1.1.0-grok45 · 18664 in / 2886 out tokens · 46045 ms · 2026-07-10T11:55:49.827372+00:00 · methodology

0 comments
read the original abstract

White Matter Hyperintensities (WMHs) are commonly observed in brain Magnetic Resonance Imaging (MRI) scans. They are associated with various neurological conditions, including vascular and inflammatory demyelinating diseases. Despite differing in etiology, WMHs from these conditions often appear similar on Fluid Attenuated Inversion Recovery (FLAIR) images. This similarity makes differential diagnosis challenging. In this work, we highlight the potential of combining attention-based segmentation with feature-driven classification. This approach supports more accurate and efficient classification between vascular and demyelinating white matter pathologies. For segmentation, we evaluate the effectiveness of attention mechanisms, specifically the Bottleneck Attention Module (BAM) and the Convolutional Block Attention Module (CBAM). We also test different architectures, particularly Attention U-Net. In addition, we explore advanced training strategies, such as patch-based learning and a 2.5D approach, to enhance lesion detection. After segmentation, we extract morphological features from the lesion masks. We then use them to classify WMHs based on their underlying cause. Our experiments utilize five publicly available datasets with diverse imaging protocols to promote model generalizability, despite limited sample sizes. The results suggest that attention-based segmentation and feature-driven classification offer a promising direction for discriminating vascular and demyelinating white matter lesions. Further validation in larger clinical cohorts is still needed.

Figures

Figures reproduced from arXiv: 2607.08171 by Aina Tur-Serrano, Francisco J. Perales L\'opez, Gabriel Moy\`a-Alcover.

Figure 1
Figure 1. Figure 1: This figure shows the results of segmentation using 2D slices, patches [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 5 internal anchors

  1. [1]

    J. M. Wardlaw, M. C. Valdés Hernández, S. Muñoz-Maniega, What are white matter hyperintensities made of? relevance to vascular cognitive impairment, Journal of the American Heart Association 4 (6) (2015) e001140, erratum in: J Am Heart Assoc. 2016 Jan 13;5(1):e002006. doi: 10.1161/JAHA.115.002006. doi:10.1161/JAHA.114.001140

    work page doi:10.1161/jaha.115.002006 2015

  2. [2]

    L. J. Zhang, D. C. Tian, L. Yang, K. Shi, Y . Liu, Y . Wang, F. D. Shi, White matter disease derived from vascular and demyelinating origins, Stroke and Vascular Neurology 9 (4) (2024) 344–350. doi:10.1136/svn- 2023-002791

    work page doi:10.1136/svn- 2024

  3. [3]

    Attention U-Net: Learning Where to Look for the Pancreas

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Mis- awa, K. Mori, S. McDonagh, N. Y . Hammerla, B. Kainz, B. Glocker, D. Rueckert, Attention u-net: Learning where to look for the pancreas (2018). arXiv:1804.03999. URLhttps://arxiv.org/abs/1804.03999

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  4. [4]

    J. Park, S. Woo, J.-Y . Lee, I. S. Kweon, Bam: Bottleneck attention module (2018). arXiv:1807.06514. URLhttps://arxiv.org/abs/1807.06514

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    S. Woo, J. Park, J.-Y . Lee, I. S. Kweon, Cbam: Convolutional block at- tention module (2018). arXiv:1807.06521. URLhttps://arxiv.org/abs/1807.06521

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    H. Ocal, Dsbav-net: Depthwise separable bottleneck attention v- shaped network with hybrid convolution for left atrium segmentation, Arabian Journal for Science and Engineering 50 (2025) 1097–1108. doi:10.1007/s13369-024-09131-1. URLhttps://doi.org/10.1007/s13369-024-09131-1

    work page doi:10.1007/s13369-024-09131-1 2025

  7. [7]

    Zhang, Z

    Y .-D. Zhang, Z. Zhang, X. Zhang, S.-H. Wang, Midcan: A multiple in- put deep convolutional attention network for covid-19 diagnosis based on chest ct and chest x-ray, Pattern Recognition Letters 150 (2021) 8–16. doi:10.1016/j.patrec.2021.06.021. URLhttps://www.sciencedirect.com/science/article/pii/ S0167865521002270

    work page doi:10.1016/j.patrec.2021.06.021 2021

  8. [8]

    G. Ali, M. M. Eid, O. G. Ahmed, M. Abotaleb, A. M. Z. Alaabdin, B. A. Buruga, Artificial intelligence in corneal topography: A short article in enhancing eye care, Mesopotamian Journal of Artificial Intelligence in Healthcare 2023 (2023) 31–34. doi:10.58496/MJAIH/2023/006. URLhttps://doi.org/10.58496/MJAIH/2023/006

    work page doi:10.58496/mjaih/2023/006 2023

  9. [9]

    Al-Quraishi, C

    T. Al-Quraishi, C. Ng, O. Mahdi, A. Gyasi, N. Al-Quraishi, Advanced ensemble classifier techniques for predicting tumor viability in osteosar- coma histological slide images, Applied Data Science and Analysis (2024) 52–68doi:https://doi.org/10.58496/ADSA/2024/006

    work page doi:10.58496/adsa/2024/006 2024

  10. [10]

    Kuijf, M

    H. Kuijf, M. Biesbroek, J. de Bresser, R. Heinen, C. Chen, W. van der Flier, F. Barkhof, M. Viergever, G. Biessels, Data of the white matter hyperintensity (wmh) segmentation challenge [dataset], DataverseNL,v1 (2022). doi:10.34894/AECRSD. URLhttps://doi.org/10.34894/AECRSD

    work page doi:10.34894/aecrsd 2022

  11. [11]

    H. J. Kuijf, M. W. M. Biesbroek, J. de Bresser, R. Heinen, C. L. H. Chen, W. M. van der Flier, F. Barkhof, M. A. Viergever, G. J. Bies- sels, Standardized assessment of automatic segmentation of white mat- ter hyperintensities and results of the wmh segmentation challenge, IEEE Transactions on Medical Imaging 38 (11) (2019) 2556–2568. doi:10.1109/TMI.2019.2905770

    work page doi:10.1109/tmi.2019.2905770 2019

  12. [12]

    Kuijf, G

    H. Kuijf, G. J. Biessels, Data of white matter hyperintensity shape and location feature analysis on brain mri; proof of principle study in patients with diabetes [dataset], DataverseNL,v1 (2023). doi:10.34894/KG5WBO. URLhttps://doi.org/10.34894/KG5WBO

    work page doi:10.34894/kg5wbo 2023

  13. [13]

    Muslim, A

    M. Muslim, A. Ali, Brain mri dataset of multiple sclerosis with consen- sus manual lesion segmentation and patient meta information [dataset], Mendeley Data, v1 (2022). doi:10.17632/8bctsm8jz7.1. URLhttps://doi.org/10.17632/8bctsm8jz7.1

    work page doi:10.17632/8bctsm8jz7.1 2022

  14. [14]

    Rondinella, F

    A. Rondinella, F. Guarnera, E. Crispino, G. Russo, C. D. Lorenzo, D. Maimone, F. Pappalardo, S. Battiato, ICPR 2024 Competition on Mul- tiple Sclerosis Lesion Segmentation—Methods and Results, IEEE, 2024, p. 1–16

    work page 2024

  15. [15]

    Rondinella, E

    A. Rondinella, E. Crispino, F. Guarnera, O. Giudice, A. Ortis, G. Russo, C. D. Lorenzo, D. Maimone, F. Pappalardo, S. Bat- tiato, Boosting multiple sclerosis lesion segmentation through attention mechanism, Computers in Biology and Medicine 161 (2023) 107021. doi:10.1016/j.compbiomed.2023.107021. URLhttps://doi.org/10.1016/j.compbiomed.2023.107021

    work page doi:10.1016/j.compbiomed.2023.107021 2023

  16. [16]

    Rondinella, F

    A. Rondinella, F. Guarnera, O. Giudice, A. Ortis, G. Russo, E. Crispino, F. Pappalardo, S. Battiato, Enhancing multiple sclerosis lesion segmen- tation in multimodal mri scans with diffusion models, in: 2023 IEEE In- ternational Conference on Bioinformatics and Biomedicine Workshops (CMISF), IEEE, 2023

    work page 2023

  17. [17]

    Spreng, R

    R. Spreng, R. Setton, U. Alter, B. Cassidy, B. Darboh, E. DuPre, K. Kan- tarovich, A. Lockrow, L. Mwilambwe-Tshilobo, W.-M. Luh, P. Kundu, G. Turner, Neurocognitive aging data release with behavioral, structural, and multi-echo functional mri measures, [dataset] OpenNeuro, v1.0.13 (2022). doi:doi:10.18112/openneuro.ds003592.v1.0.13

    work page doi:10.18112/openneuro.ds003592.v1.0.13 2022

  18. [18]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, V ol. 9351 of Lecture Notes in Computer Science, Springer, 2015, pp. 234–241, available at arXiv:1505.04597 [cs.CV]. URLhttps://arxiv.org/abs/1505.04597

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  19. [19]

    Petrovi ´c, G

    N. Petrovi ´c, G. Moyà-Alcover, A. J. i Capó, M. González- Hidalgo, Sickle-cell disease diagnosis support selecting the most ap- propriate machine learning method: Towards a general and inter- pretable approach for cell morphology analysis from microscopy images, Computers in Biology and Medicine 126 (2020) 104027. doi:10.1016/j.compbiomed.2020.104027. UR...

    work page doi:10.1016/j.compbiomed.2020.104027 2020

  20. [20]

    S. M. Martorell, M. C. Blázquez, D. G. Figueredo, S. G. Ortiz, J. C. Font, Hyperintense punctiform images in the white matter: A diag- nostic approach, Radiología (English Edition) 54 (4) (2012) 321–335. doi:10.1016/j.rxeng.2011.09.001. URLhttps://doi.org/10.1016/j.rxeng.2011.09.001

    work page doi:10.1016/j.rxeng.2011.09.001 2012

  21. [21]

    2.5D Deep Learning for CT Image Reconstruction using a Multi-GPU implementation

    A. Ziabari, D. H. Ye, S. Srivastava, K. D. Sauer, J.-B. Thibault, C. A. Bouman, 2.5d deep learning for ct image reconstruction using a multi-gpu implementation, arXiv preprint arXiv:1812.08367 (2018). arXiv:1812.08367. URLhttps://arxiv.org/abs/1812.08367

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  22. [22]

    Shin, Revolutionizing medical imaging with artificial intelligence: Real-time segmentation for enhanced diagnostics, EDRAAK (2024) 18– 25doi:10.70470/EDRAAK/2024/003

    J. Shin, Revolutionizing medical imaging with artificial intelligence: Real-time segmentation for enhanced diagnostics, EDRAAK (2024) 18– 25doi:10.70470/EDRAAK/2024/003

    work page doi:10.70470/edraak/2024/003 2024

  23. [23]

    Y . Yang, H. Wang, C. Ji, Y . Niu, Artificial intelligence-driven diag- nostic systems for early detection of diabetic retinopathy: Integrat- ing retinal imaging and clinical data, SHIFAA 2023 (2023) 83–90. doi:10.70470/SHIFAA/2023/010

    work page doi:10.70470/shifaa/2023/010 2023

  24. [24]

    Sheela, G

    M. Sheela, G. Amirthayogam, J. J. Hephzipah, R. Suganthi, T. Karthikeyan, M. Gopianand, Advanced brain tumor classification using deepbeleif-cnn method, Babylonian Journal of Machine Learning 2024 (2024) 89–101. doi:10.58496/BJML/2024/009. URLhttps://doi.org/10.58496/BJML/2024/009

    work page doi:10.58496/bjml/2024/009 2024

  25. [25]

    H. I. W. Al-Shahwani, A. K. Faieq, The benefit of artificial intel- ligence in the analysis of malignant brain diseases: A mini review, Mesopotamian Journal of Artificial Intelligence in Healthcare (2023) 57– 60doi:10.58496/MJAIH/2023/011

    work page doi:10.58496/mjaih/2023/011 2023

  26. [26]

    M. A. Westenberg, J. B. T. M. Roerdink, M. H. F. Wilkinson, V olumetric attribute filtering and interactive visualization using the max-tree repre- sentation, IEEE Transactions on Image Processing 16 (12) (2007) 2943–

    work page 2007

  27. [27]

    doi:10.1109/TIP.2007.909317

    work page doi:10.1109/tip.2007.909317 2007

  28. [28]

    X. Chen, Y . Li, L. Yao, E. Adeli, Y . Zhang, X. Wang, Generative adversar- ial u-net for domain-free few-shot medical diagnosis, Pattern Recognition Letters 157 (2022) 112–118. doi:10.1016/j.patrec.2022.03.022. URLhttps://www.sciencedirect.com/science/article/pii/ S0167865522000873

    work page doi:10.1016/j.patrec.2022.03.022 2022

  29. [29]

    Z. Yang, Z. Zhao, Y . Gu, Y . Xu, Query-guided generalizable medi- cal image segmentation, Pattern Recognition Letters 184 (2024) 52–58. doi:https://doi.org/10.1016/j.patrec.2024.06.005

    work page doi:10.1016/j.patrec.2024.06.005 2024