LV-Net: Anatomy-aware lateral ventricle shape modeling with a case study on Alzheimer's disease

Jinah Park (for the Alzheimer's Disease Neuroimaging Initiative; Lifestyle flagship study of ageing); Suhyun Ahn; the Australian Imaging Biomarkers; Wonjung Park

REVIEW 4 major objections 5 minor 3 references

LV-Net: Anatomy-aware lateral ventricle shape modeling with a case study on Alzheimer's disease

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read LV-Net reconstructs whole lateral ventricle surfaces from brain MRI by deforming a joint ventricle-hippocampus template, and uses the stable vertex correspondence to localize Alzheimer's-associated ventricle subregions.

desk verdict Solid anatomical prior, under-specified core module, and no error bars; worth a serious referee but with mandatory revisions. read the letter →

arxiv 2508.06055 v2 pith:6HJFDEJR submitted 2025-08-08 cs.CV cs.GR

Wonjung Park , Suhyun Ahn , Jinah Park (for the Alzheimer's Disease Neuroimaging Initiative , the Australian Imaging Biomarkers , Lifestyle flagship study of ageing) This is my paper · ORCID

classification cs.CVcs.GR

keywords lateralventricleshapemodelingdeformabletemplatemeshpointcorrespondenceAlzheimer'sdiseaseMRIsegmentationstatisticalanalysisdeeplearning

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

The lateral ventricle is the C-shaped fluid-filled cavity at the center of the brain; its shape is a candidate biomarker for Alzheimer's and other neurological diseases, but automatic modeling has been blocked by large normal variation and by MRI segmentations that fragment or miss the thin inferior part of the ventricle. The paper's claim is that an anatomy-aware deformable template solves both problems: the template contains the ventricle and hippocampus as one mesh with shared boundary vertices, and every ventricle vertex carries a label saying which surrounding structure it should touch. The shared boundary lets the better-segmented hippocampus guide the ventricle where its own mask is broken, and the labels force corresponding vertices to land on the same anatomical structure in every subject. On three public brain MRI datasets, LV-Net reports better Dice, average surface distance, and 95th-percentile Hausdorff scores than registration-based and particle-correspondence baselines, plus a more compact and generalizable shape space. The payoff is a practical route from imperfect clinical MRI to whole-ventricle meshes whose vertex positions can be compared statistically, which the paper uses to identify ventricle subregions associated with Alzheimer's disease.

What carries the argument

The anatomy-aware joint LV-hippocampus template mesh is a deformable mesh pair that shares a strip of boundary vertices between the ventricle and the hippocampus, with each ventricle vertex labeled by one of five surrounding structures. This template is the load-bearing object: its fixed topology supplies the parameterization, its shared boundary gives the hippocampus leverage to repair missing inferior ventricle segments, and its per-vertex labels define what correspondence means. The fitting mechanism is an iterative loop in which a neural vertex-deformation module predicts displacements, a distance loss (Chamfer distance plus point-to-mesh and mesh-to-point terms) pulls the mesh to the ta

What would settle it

Run the paper's own ablation quantitatively: take a set of low-resolution segmentations with inferior-LV holes, reconstruct with the full joint template and with the paper's LV-Net* variant, and compare both to a high-resolution reference using ASSD restricted to the inferior horn. If the joint template does not beat the LV-Net* variant by more than its internal variability, the anatomical-correction claim has no support. A sharper version: rerun the 5,000-iteration optimization with a randomly initialized vertex-deformation module; since all deformation comes from that module's predicted disp

Watch

Extended reading notes

Core claim

The central claim is that anatomical prior knowledge, encoded in the template rather than applied afterward, is what makes LV shape reconstruction robust. LV-Net begins with a joint LV-hippocampus template mesh whose boundary vertices belong to both structures; during iterative deformation, the hippocampus mesh is fitted to its own target point cloud, and the shared vertices pull the inferior ventricle into anatomically correct position even where the ventricle mask has holes. Each LV vertex is assigned one of five periventricular labels, and a per-region distance loss keeps labeled vertices near same-labeled target points, giving point correspondence that follows anatomy rather than raw geo

Load-bearing premise

The whole optimization depends on a neural network that predicts vertex displacements at every step, but the paper never states how that network is trained, what data it uses, or how it is initialized; if that module is unavailable or under-trained, the iterative template deformation cannot be reproduced as written.

Editorial extensions

If this is right

Whole-ventricle analysis becomes feasible in routine datasets: studies no longer need to discard the inferior horn because its mask is disconnected or absent.
Automatic segmentation outputs with holes or fragments are still usable, because the hippocampus-anchored template restores the missing boundary anatomy.
Vertex-level shape statistics are interpretable across subjects, since corresponding vertices are tied to the same periventricular structure rather than to geometric nearest neighbors.
The Alzheimer's case study pinpoints candidate local biomarkers: ventricle subregions adjacent to the hippocampus, thalamus, and part of the caudate are where shape distributions differ between AD and cognitively normal groups.

Reading between the lines

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

Beyond the paper, the shared-boundary repair mechanism should transfer to other anatomical pairs where one structure is harder to segment than a neighboring structure; each pair would need its own template, but the repair logic is generic.
A direct test the paper does not run: feed LV-Net target point clouds from a second segmentation method and see whether the reported accuracy gap over the baselines persists; the current experiments hinge on point clouds from a single automatic segmenter.
The AD map is built from per-vertex Wilcoxon tests at $p \le 0.1$ with no stated multiple-comparison correction; a permutation test with false-discovery-rate control would show which parts of the map, especially the caudate-adjacent region, are robust.
Because the template fixes topology, the fitted meshes could be subdivided anatomically (frontal horn, body, atrium, temporal horn) for subregional shape analysis, a level of localization the current vertex-wise map only hints at.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Colleague,

Here's my read on LV-Net. The genuinely new thing is the joint LV-hippocampus template with shared boundary vertices and periventricular vertex labels. Using hippocampal adjacency to constrain the inferior horn, which is often fragmented in segmentations, is a sensible anatomical prior, and the qualitative results show it working—LV-Net reconstructs the inferior horn where ShapeWorks and LDDMM produce folded or incomplete shapes. That's a real contribution to LV shape modeling, not a new paradigm, but a useful tool for the community.

What the paper does well: three datasets, two baseline methods, and a set of statistical shape metrics (generalization, specificity, compactness) in addition to the usual DSC/ASSD/HD95. The improvement over baselines is consistent and plausible. The AD case study finds significant shape differences adjacent to the hippocampus, which matches prior volumetric findings; that's a good sanity check.

The soft spots are mostly about evidence, not about the core idea. The biggest is the PointNet-based vertex deformation module in §3.2.2. The paper never says what the network inputs are, its architecture, how it's trained, or whether it's pretrained and frozen or optimized jointly with the mesh vertices during the 5,000 AdamW iterations. That's load-bearing: every reconstruction in Table 1 goes through that module, so the method as written is not implementable. The GitHub link may resolve it, but the text should stand alone. This is a major fix, not a nicety.

Second, Table 1 reports aggregate means without error bars or significance tests. The claim "outperforms all methods" is too strong without that. Third, the "error correction" ablation compares low-resolution reconstructions against high-resolution reconstructions, both produced by LV-Net. It shows the joint template is robust to input resolution, which is fine, but it's not a test against an independent ground truth. The AD analysis uses p ≤ 0.1 without multiple-comparison correction; that's acceptable for an exploratory visualization, but they should say so.

Overall, the central argument holds up. The anatomical prior is sound, and the qualitative evidence supports it. The paper needs a referee but also needs revision: specify the PointNet, add error bars/statistics, and temper the correction claim. I'd send it to peer review, but conditional on those changes.

Serious thinker: yes. The work is coherent and honestly framed; the weaknesses are in evaluation transparency, not in reasoning.

I'd bring it to a specialist reading group, though maybe not a general one. I'd cite it if I work on LV shape, once the module is specified.

Referee Report

4 major / 5 minor

Summary. LV-Net proposes an automated pipeline for reconstructing subject-specific lateral ventricle (LV) meshes from brain MRI. The method deforms an anatomy-aware LV–hippocampus joint template mesh toward a target point cloud extracted from SynthSeg segmentations. The template includes shared boundary vertices between the LV and hippocampus and per-vertex labels for adjacent subcortical structures. Reconstruction is cast as 5,000 AdamW iterations of mesh deformation driven by the distance and regularization losses in Eqs. (1) and (2), with a PointNet-based vertex deformation module mentioned as the displacement predictor. Experiments compare LV-Net with LDDMM and ShapeWorks on OASIS, ADNI, and AIBL using DSC, ASSD, and HD95, and also evaluate statistical shape model quality (generalization, specificity, compactness). A case study applies the method to detect LV subregions associated with Alzheimer's disease using vertex-wise Wilcoxon tests.

Significance. The core idea of jointly modeling the LV and hippocampus to compensate for segmentation loss in the inferior LV, and of embedding periventricular anatomical labels into the template for correspondence, is original and clinically plausible. The code release is a clear strength. If the implementation were fully specified and the quantitative claims properly supported, LV-Net could be a useful contribution to anatomical shape modeling. As it stands, however, the central quantitative results rest on an underspecified deformation module and on aggregate statistics without variance or significance testing, and the AD analysis uses a lenient uncorrected threshold. These issues currently limit the reliability of the stated conclusions.

major comments (4)

[§3.2.2] The core deformation step is not implementable as written. The text states only that 'a PointNet-based vertex deformation module (Qi et al., 2017) predicts vertex displacements' but gives no architecture, input features, layer sizes, parameter count, training data, or initialization. It is also unclear whether the module is pretrained, frozen, or optimized jointly during the 5,000 AdamW iterations. Since every subject-specific reconstruction in Table 1 is produced by this module, the claimed superiority over LDDMM and ShapeWorks cannot be reproduced or falsified from the manuscript. Please provide a complete specification, or correct the description if the module is not a learned network.
[Table 1 / §4.2] The central comparison reports only three aggregate numbers (DSC 0.893, ASSD 0.715 mm, HD95 2.064 mm for LV-Net), with no per-dataset breakdown, no subject counts, no standard deviations/confidence intervals, and no significance tests. The statement that 'LV-Net outperforms all methods' is therefore supported only by point estimates. The table title also promises 'ASSD for each subcortical region' but the table has no such columns. Please provide per-dataset results with error bars and appropriate statistical tests, and repair the table/hierarchy mismatch.
[§5, Figure 7] The AD subregion analysis uses a vertex-wise Wilcoxon signed-rank test at p≤0.1 with no multiple-comparison correction. With 2,490 LV vertices, even under a global null one would expect roughly 249 vertices to fall below this threshold by chance, so the reported 'significant' regions are not a reliable basis for clinical claims. In addition, the signed-rank test is normally for paired observations, whereas the comparison is between independent CN and AD groups; a Mann-Whitney U test or a model adjusted for age/sex would be more appropriate. Please apply a corrected threshold or permutation-based FDR control and use a standard significance level.
[§4.3, Figure 4(a)] The error-correction validation compares LV-Net reconstruction from a downsampled mask with LV-Net reconstruction from the high-resolution mask. This is a self-consistency check: it shows the method is stable under resolution changes, but it does not establish that the restored inferior LV is anatomically correct, because both reconstructions come from the same method and share the same template bias. The claim that the joint template 'effectively restores the inferior LV' needs an independent reference, such as the original high-resolution segmentation or manual ground truth, and should be quantified rather than shown with a single qualitative example.

minor comments (5)

[§3.1 and §3.2.2] Section 3.1 lists five peripheral brain structures, including the hippocampus, but Section 3.2.2 defines only four peripheral regions (m=4) and omits the hippocampus from the peripheral loss. Please clarify this taxonomy and state explicitly how the hippocampal label is handled.
[Table 1] LV-Net* is discussed in the text as an important ablation for the joint template, but it does not appear in the quantitative table. Adding it to Table 1 (or a companion table) would make the ablation quantitative.
[Figure 5] The generalization/specificity/compactness plots are described as supporting 'higher or comparable' performance, but no error bars or subject-level variability are shown and the legend is not described. Please add variability information.
[General] Typos and small errors: 'Housdroff' should be 'Hausdorff'; 'significantly associations' should be 'significant associations'; the reference 'Shaphiro, S.' should be 'Shapiro, S.'; and 'The codes' is nonstandard ('The code').
[§3.1] The paper does not describe how the joint LV-hippocampus template mesh is constructed (e.g., which atlas, manual labeling, or prior method). Since the template is central to the method, please provide this information.

Circularity Check

1 steps flagged · score 3.0 of 10

One secondary validation is circular: 'error correction' is measured against LV-Net's own high-resolution reconstruction; the central reconstruction comparison is not circular.

self definitional [Section 4.3, 'Anatomy correction with the template mesh' (Figure 4a)]
"We then reconstructed the low-resolution masks with and without the joint template mesh and compared them to the high-resolution reconstruction. As illustrated in Figure 4(a), the joint template mesh effectively restores the inferior LV to a shape nearly identical to the high-resolution reconstruction, demonstrating its anatomical correction capability."

The 'high-resolution reconstruction' used as the reference for evaluating error correction is itself produced by LV-Net. Therefore, the experiment compares two outputs of the same model: one fitted to a low-resolution mask and one fitted to a high-resolution mask. Agreement between these two outputs shows self-consistency of LV-Net, not that the restored inferior LV matches true anatomy. The conclusion 'anatomical correction capability' is thus supported by a reference that is generated by the method under evaluation, rather than by an independent ground truth. This makes the validation circular for the specific claim of anatomical correction, although it does not undermine the main reconstruction comparison against LDDMM and ShapeWorks.

full rationale

The central derivation of LV-Net is a per-subject fitting procedure: a fixed joint template is deformed to target point clouds extracted from segmentation maps, with losses that directly minimize geometric distance to those targets. The headline reconstruction metrics (DSC, ASSD, HD95) evaluate how well each method fits the same target geometry, so the comparison against LDDMM and ShapeWorks is a fair benchmark rather than a circular prediction. The peripheral-region loss is an explicit design term that enforces correspondence by anatomical label; it is a construction choice, not a circular derivation. The only circular element found is the error-correction validation in Section 4.3, which uses LV-Net's own high-resolution reconstruction as the reference for 'anatomical correction.' This does not affect the main claim's independence but does weaken the secondary claim. The under-specification of the PointNet-based vertex deformation module is a reproducibility concern, not a circularity, since no equation is defined in terms of the claimed result and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the joint template being anatomically correct, the PointNet module being functional, and SynthSeg segmentation being accurate. Many hyperparameters are hand-tuned; all are listed in the ledger. No new physical entities are introduced.

free parameters (6)

Distance loss weights λ_cf, λ_pm, λ_mp = 2, 1.4, 0.6
Hand-set weights balancing Chamfer, point-to-mesh, and mesh-to-point distances; no sensitivity analysis.
Regularization weights λ_vert, λ_norm, λ_edge, λ_cons, λ_lap = 1, 1, 1000, 100, 300
Hand-set to enforce smooth, plausible deformation; no ablation on these values.
Template resolution = 2 mm edge, 2,490 LV vertices, 96 shared vertices
Template mesh resolution is a design choice affecting fidelity and computational cost.
Optimization schedule = 5000 iterations, lr 5e-4 halved every 1000 iterations
Stopping criterion and learning rate are chosen by hand.
Peripheral region count m = 4
Number of peripheral anatomical regions used in the peri-distance term; excludes hippocampus as separate loss.
AD significance threshold = p <= 0.1
Threshold for vertex-wise Wilcoxon tests in the AD case study; lenient and no multiple-comparison correction.

assumptions (4)

domain assumption The PointNet-based vertex deformation module is available and predicts useful vertex displacements.
Section 3.2.2 states the module predicts displacements but never specifies its architecture, training data, or optimization; the whole pipeline depends on it.
domain assumption The LV-hippocampus joint template mesh and its periventricular vertex labels are anatomically correct.
Section 3.1 constructs the template based on anatomical adjacency; errors in the template would propagate to every reconstruction.
domain assumption SynthSeg segmentation maps provide sufficiently accurate target point clouds and peripheral labels.
Section 3.2.1 uses SynthSeg outputs as the only source of target geometry; segmentation errors become target errors.
ad hoc to paper Shared vertices along the LV-hippocampus boundary enforce correct inferior LV structure when LV masks are incomplete.
This is the paper's core design bet (Section 3.1); it is validated only by self-comparison of LV-Net outputs at different resolutions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LV-Net: Anatomy-aware lateral ventricle shape modeling with a case study on Alzheimer's disease." pith.science (2026). https://pith.science/paper/6HJFDEJR

@misc{pith2026250806055,
  author       = {Pith},
  title        = {Pith review of: LV-Net: Anatomy-aware lateral ventricle shape modeling with a case study on Alzheimer's disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6HJFDEJR}},
  note         = {Machine review of arXiv:2508.06055}
}

read the original abstract

Lateral ventricle (LV) shape analysis holds promise as a biomarker for neurological diseases; however, challenges remain due to substantial shape variability across individuals and segmentation difficulties arising from limited MRI resolution. We introduce LV-Net, a novel framework for producing individualized 3D LV meshes from brain MRI by deforming an anatomy-aware joint LV-hippocampus template mesh. By incorporating anatomical relationships embedded within the joint template, LV-Net reduces boundary segmentation artifacts and improves reconstruction robustness. In addition, by classifying the vertices of the template mesh based on their anatomical adjacency, our method enhances point correspondence across subjects, leading to more accurate LV shape statistics. We demonstrate that LV-Net achieves superior reconstruction accuracy, even in the presence of segmentation imperfections, and delivers more reliable shape descriptors across diverse datasets. Finally, we apply LV-Net to Alzheimer's disease analysis, identifying LV subregions that show significantly associations with the disease relative to cognitively normal controls. The codes for LV shape modeling are available at https://github.com/PWonjung/LV_Shape_Modeling.

Figures

Figures reproduced from arXiv: 2508.06055 by the authors.

**Figure 1.** (a) LV-hippocampus joint template mesh with shared vertices between LV and hippocampus and (b) anatomy-aware joint template mesh with adjacent area information. First, to preserve anatomical correctness, we introduce the LV-hippocampus joint template mesh of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Overall framework to model the target LV and hippocampus mesh from the joint template mesh. 𝑚𝑝, accounting for missing points in the target point cloud due to segmentation errors. The regularization loss ensures plausible mesh deformation and is derived from each deformed LV and hippocampus meshes. The regularization loss is : 𝑟𝑒𝑔 =  𝐿 𝑟𝑒𝑔 +  ℎ 𝑟𝑒𝑔 (2) where  𝑖 𝑟𝑒𝑔 =𝜆𝑣𝑒𝑟𝑡||Δ𝑣𝑒𝑟𝑡𝑖 ||2 + 𝜆𝑛𝑜𝑟𝑚||Δ𝑛𝑜𝑟𝑚𝑖 ||2 + 𝜆𝑒𝑑𝑔… view at source ↗

**Figure 3.** Examples of optimization results according to the shape modeling methods. Red points are target shapes on the first row of each subject. The colors of spectral particles indicate correspondence for each method. LV-Net* indicates that when using only the LV template mesh 𝑀𝑙 along with periventricular information [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: (a) Error correction by joint template mesh-based LV reconstruction. The inferior LV is accurately restored to match the high-resolution shape. (b) Guiding template vertices to corresponding subcortical areas using peripheral region information yields more reasonable r…

**Figure 5.** Figure 5: Quantitative plots of generalization, specificity and compactness on normal and demented groups of OASIS, ADNI, and AIBL datasets with 50 subjects each. uses LV template mesh, the absence of guidance due to missing LV masks induces skewed mesh artifacts (i.e. nonunifo…

**Figure 6.** Figure 6: Visualization of the first 3 PCA components of the learned LV shape spaces. We modulated the shape in each dimension using 3𝜎 and -3𝜎 variations. L R p -value 0 ≥ 0.1 R L [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Visualization of the lateral ventricle indicating subregions where vertex-wise shape distributions significantly differ between Alzheimer’s disease subjects and cognitively normal controls. (superior and inferior views) Alzheimer’s disease (AD). In addition, the portio…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 1 canonical work pages

[2008]

Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 27, 685–691

The alzheimer’s disease neuroimaging initiative (adni): Mri methods. Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 27, 685–691. Johnson, S.L., Wang, L., Alpert, K.I., Greenstein, D., Clasen, L., Lalonde, F., Miller, R., Rapoport, J., Gogtay, N., 2013. Hippocampal shape abnormalit...

arXiv 2013
[2021]

Deepcsr: A 3d deep learning approach for cortical surface reconstruction,in:ProceedingsoftheIEEE/CVFWinterConferenceon Applications of Computer Vision, pp. 806–815. Desbrun, M., Meyer, M., Schröder, P., Barr, A.H., 1999. Implicit fairing of irregular meshes using diffusion and curvature flow, in: Proceedings of the 26th annual conference on Computer graph...

arXiv 1999
[2022]

IEEE Transactions on Medical Imaging 42, 430–443

Cortexode: Learning cortical surface reconstruction by neural odes. IEEE Transactions on Medical Imaging 42, 430–443. Marcus, D.S., Fotenos, A.F., Csernansky, J.G., Morris, J.C., Buckner, R.L., 2010. Open access series of imaging studies: longitudinal mri data in nondemented and demented older adults. Journal of cognitive neuroscience 22, 2677–2684. Olive...

work page 2010

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [2008]

Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 27, 685–691

The alzheimer’s disease neuroimaging initiative (adni): Mri methods. Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 27, 685–691. Johnson, S.L., Wang, L., Alpert, K.I., Greenstein, D., Clasen, L., Lalonde, F., Miller, R., Rapoport, J., Gogtay, N., 2013. Hippocampal shape abnormalit...

arXiv 2013

[2] [2021]

Deepcsr: A 3d deep learning approach for cortical surface reconstruction,in:ProceedingsoftheIEEE/CVFWinterConferenceon Applications of Computer Vision, pp. 806–815. Desbrun, M., Meyer, M., Schröder, P., Barr, A.H., 1999. Implicit fairing of irregular meshes using diffusion and curvature flow, in: Proceedings of the 26th annual conference on Computer graph...

arXiv 1999

[3] [2022]

IEEE Transactions on Medical Imaging 42, 430–443

Cortexode: Learning cortical surface reconstruction by neural odes. IEEE Transactions on Medical Imaging 42, 430–443. Marcus, D.S., Fotenos, A.F., Csernansky, J.G., Morris, J.C., Buckner, R.L., 2010. Open access series of imaging studies: longitudinal mri data in nondemented and demented older adults. Journal of cognitive neuroscience 22, 2677–2684. Olive...

work page 2010