Pith. sign in

REVIEW 4 major objections 5 minor 17 references

Semantic Segmentation for Preoperative Planning in Transcatheter Aortic Valve Replacement

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Focal skeleton recall loss plus rule-generated pseudo-labels lifts TAVR anatomy segmentation to 83.2% Dice.

desk verdict A useful public dataset and a credible loss tweak, but the pseudo-label heuristics are unvalidated, so treat the benchmark numbers as measuring the heuristic labels, not anatomy. read the letter →

arxiv 2507.16573 v1 pith:2LMX3IIO submitted 2025-07-22 eess.IV cs.CV

Cedric Zöllner , Simon Rei{ss} , Alexander Jaus , Amroalalaa Sholi , Ralf Sodian , Rainer Stiefelhagen This is my paper · ORCID
classification eess.IVcs.CV
keywords semanticsegmentationTAVRplanningtranscatheteraorticvalvereplacementpseudo-labelsfocallossskeletonrecallcomputedtomographyroot
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 sets out to make transcatheter aortic valve replacement (TAVR) planning measurable by adding the missing anatomical structures — aortic root, valve, and annulus — to an existing public CT dataset using rule-based geometric heuristics, and by training a single semantic segmentation model to find all seven planning-relevant structures. To improve learning on thin, tubular, and underrepresented structures, the authors propose a focal skeleton recall loss that weights skeleton voxels by prediction confidence. On the enriched test set, the loss lifts Swin UNETR's mean Dice — a standard overlap metric between prediction and ground truth — from 81.86% to 83.20% and mean IoU from 72.86% to 74.79%, with the largest gains on the left and right iliac arteries. If correct, the work provides a public benchmark and a route to automatic TAVR measurements from CT without additional expert annotations for these structures.

What carries the argument

The central new mechanism is the focal skeleton recall loss (Eq. 3), a modification of the skeleton recall loss in which each voxel on the thinned class skeleton is weighted by $(1-p_{i,c})^\gamma$, so low-confidence predictions on thin structures are emphasized dynamically during training; it is combined with the focal loss as an auxiliary objective. The other load-bearing component is the pseudo-label enrichment pipeline: distance thresholds (three voxels for the valve, one voxel for the annulus) and a cross-sectional-area local-minimum search (with a moving average) define the aortic root end. The loss carries the performance claim; the heuristics carry the dataset claim.

What would settle it

Have an expert manually segment the valve, annulus, and aortic root on a random sample of the test CTs and compute Dice between those expert labels and the rule-based pseudo-labels; if agreement on the annulus (reported model Dice ~55%) is far below the other classes, the benchmark mostly measures how well models reproduce the label-generation rules. Alternatively, compare automated measurements (annulus diameter, root height, iliac artery diameter) against the values recorded in real TAVR planning workflows.

Watch

Extended reading notes

Core claim

The paper claims that the TAVR-relevant anatomy absent from public datasets can be synthesized from coarse anatomy labels by three geometric rules: the valve is the aortic voxels within three voxels of the left ventricle, the annulus is the left-ventricle voxels within one voxel of the aorta, and the aortic root is the aortic segment from the annulus to the first local minimum of the aorta's cross-sectional area. A segmentation network trained on the resulting 578-scan, seven-class dataset can locate these structures, and the paper's focal skeleton recall loss improves that capability: with Swin UNETR, mean Dice on the test set goes from 81.86% (Dice plus cross-entropy loss) to 83.20%, and mean IoU from 72.86% to 74.79%. The paper presents this as a step toward automatic measurement of annulus size, root dimensions, and vascular access from CT in TAVR preoperative planning.

Load-bearing premise

The load-bearing premise is that the automatically generated labels for the aortic root, valve, and annulus, produced by voxel-distance thresholds and a cross-sectional-area minimum, match the anatomical definitions surgeons actually use in TAVR planning; if they do not, the trained models are learning the label-generation rules rather than the true anatomy, and the benchmark scores overstate clinical utility.

Editorial extensions

If this is right

  • The enriched dataset of 578 CTs with seven TAVR-relevant classes becomes a public benchmark, so other groups can train and compare segmentation models without in-house expert annotations.
  • A single model reaches 83.2% mean Dice across all seven structures, so annulus size, aortic root dimensions, and iliac artery diameters can be derived from the segmentations automatically rather than by hand tracing.
  • The focal skeleton recall loss improves connectedness of thin tubular structures (iliac artery Dice +1.33/+1.72 points) and removes disconnected false segments, and this benefit is expected for any segmentation task with thin structures.
  • Validation experiments show the proposed loss outperforms both the Dice-plus-cross-entropy baseline (+1.72% Dice) and focal loss alone (+0.42% Dice), so the gain is not just a baseline artifact.
  • Raising the mean IoU by +1.93 points alongside Dice indicates the model's predictions overlap the target labels spatially, not merely at the surface.

Reading between the lines

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

  • Because the valve, annulus, and aortic root labels come from geometric rules rather than expert annotation, the reported Dice scores measure agreement with those rules; clinical deployment would first require checking the rules against manual expert segmentations or actual TAVR measurements.
  • The same confidence-weighted skeleton-loss idea could plausibly transfer to other tubular anatomies — coronary arteries, carotid vessels, bronchi — where connectivity matters and annotations are scarce.
  • A natural extension is adding the structures the paper names as still missing (calcifications, femoral arteries, coronary ostia) and testing whether the heuristic labels for those can be generated from existing calcium-scoring or atlas methods.
  • The aortic root end defined by the first cross-sectional-area minimum is a property of the whole dataset shown in the aggregate; per-scan variability in that heuristic is untested, and a learned regressor for the root end could be compared against it.
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.

Referee Report

4 major / 5 minor

Summary. This paper addresses semantic segmentation for TAVR preoperative planning. The authors start from TotalSegmentator coarse annotations (aorta, left ventricle, iliac arteries) and derive fine-grained pseudo-labels for aortic root, valve, and annulus via geometric heuristics (voxel-distance thresholds and cross-sectional-area minima). They release the enriched dataset publicly. They then propose a focal skeleton recall loss (Eq. 3), combining focal weighting with skeleton recall, and evaluate U-Net, V-Net, and Swin UNETR on a 378/100/100 split. The best configuration, Swin UNETR with the proposed loss, reaches 83.20% mean Dice on the test set, a +1.34 percentage-point improvement over DiceCE.

Significance. The public release of enriched pseudo-labels is a useful community resource if the labels are anatomically valid, and the proposed loss is simple and generally applicable. The paper is honest about limitations (Section 4). However, because all evaluation is against unvalidated pseudo-labels and the loss comparison lacks error bars, the central quantitative claims should be treated as provisional.

major comments (4)
  1. [Section 2.1] The valve, annulus, and aortic root pseudo-labels are defined by geometric heuristics — valve as aorta voxels within 3 voxels of the left ventricle, annulus as left-ventricle voxels within 1 voxel of the aorta, and aortic root end as the first local minimum of cross-sectional area after the first local maximum — rather than by the anatomical definitions used in TAVR planning (e.g., the annulus is the virtual basal ring at the leaflet hinge points). No expert manual annotations are used to validate these derived labels. Since every result in Tables 2 and 3 is computed against these pseudo-labels, a systematic mismatch between the heuristics and true TAVR anatomy would make the benchmark measure agreement with the label-generation rules rather than clinically meaningful segmentation. This is load-bearing for both the dataset contribution and the loss claim; at minimum, the authors should provide expert visual validation of a random subset, quantitative comparison with manual TAVR measurements on a small set, or an explicit statement of the anatomical proxies and their known limitations.
  2. [Section 3, Tables 2 and 3] All comparisons use a single fixed train/validation/test split and one training run per configuration. The reported differences are small (e.g., +1.34% Dice for Swin UNETR, but −1.30% for V-Net with the same loss change), and without standard deviations across seeds or a paired significance test it is not possible to conclude that the focal skeleton recall loss reliably improves performance, especially for the small classes (annulus, valve). The authors should report mean and standard deviation over at least 3–5 seeds and perform a paired test (e.g., Wilcoxon or bootstrap) on the per-case Dice scores.
  3. [Abstract and Section 4] The abstract states '+1.27% Dice increase', while Section 4 and Table 3 report +1.34% for Swin UNETR on the test set and Table 2 reports +1.72% on the validation set. These numbers should be reconciled and the reference point (validation vs test, DiceCE vs focal loss) stated consistently.
  4. [Section 2.1, Figure 1] The description of the aortic root detection is ambiguous: Figure 1 plots the cross-sectional area 'over the entire dataset' and reports the local minimum at distance 25, which suggests a dataset-level constant, while the text says the end is identified individually in each CT. If the same distance is applied to all patients, the aortic root label is not patient-specific and may be systematically wrong for patients with different aortic geometries. Please clarify whether the detection is per-CT and, if a pooled curve is used only for illustration, state this explicitly. Also specify how the moving average window is chosen.
minor comments (5)
  1. [Section 2.2] The notation is inconsistent: Equation (3) writes LFocalSR, but Section 3 and Table 2 use LFocalSK and LFocalSK⋆. Please unify the notation.
  2. [Section 2.1] The distance thresholds for valve and annulus are given in voxels, but CT voxel spacings vary across scans. Specify whether the thresholds are in physical units or voxels and, if voxels, whether anisotropic spacing was accounted for.
  3. [Section 3] The sentence 'we train each model with a varying batch size of CT scans' is unclear; list the effective batch size (number of subvolumes per update) for each architecture explicitly.
  4. [Table 1] Please describe how the 578 CTs were selected from the 1,228 TotalSegmentator scans; the current text only says some CTs 'do not show TAVR-relevant anatomy.'
  5. [Figure 2] The caption uses 'ourLFocalSK' without the star superscript and also calls it 'skeleton focal loss' in the text below the figure; align the caption with Equation (3).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: pseudo-labels are explicit rule-based transformations of external coarse labels, and the loss comparison is a standard supervised benchmark on held-out data.

full rationale

The paper's derivation chain is: TotalSegmentator coarse labels → explicit geometric rules (distance thresholds for valve/annulus, cross-section area extrema for aortic root) → fine pseudo-labels → supervised training of segmentation models → Dice/IoU evaluation on a held-out test split → loss-function comparison. No step defines its target in terms of the result it claims to predict. The pseudo-labels are rule-based operationalizations of anatomy, but the segmentation models do not receive the coarse labels or the rules at test time, so predicting them is not equivalent to applying the rules by construction. The proposed focal skeleton recall loss is a transparent modification of the cited skeleton recall and focal losses, with no fitted parameter renamed as a prediction. The +1.34% Dice gain is measured on the same pseudo-label test set for all methods, making the comparison fair even though the labels themselves lack expert validation. The only substantive weakness is clinical validity of the unvalidated pseudo-labels, which the paper itself acknowledges ('For some structures it is not yet possible to coherently segment them, such as the annulus'); this is a generalization/validation concern, not circularity. The single self-citation ([2], Jaus et al.) is used only as a dataset reference, not as load-bearing support. Therefore the derivation is self-contained and the circularity score is 0.

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

The central results rest on the unvalidated assumption that TotalSegmentator coarse labels and simple geometric rules produce anatomically correct TAVR structures, and on a training setup with several hand-picked hyperparameters (gamma, loss weights, distance thresholds). No new physical entities are introduced.

free parameters (6)
  • valve distance threshold = 3 voxels
    Chosen ad hoc to cover the curved leaflets when extracting the valve pseudo-label from aorta voxels near the left ventricle. Stated in Section 2.1.
  • annulus distance threshold = 1 voxel
    Used to define the annulus as left-ventricle voxels one voxel away from the aorta. This thin-ring definition is a heuristic choice, not validated by anatomy experts. Section 2.1.
  • aortic root cutoff = 25 voxels (from Figure 1) or local-minimum criterion
    The end of the aortic root is identified as the first local minimum of aortic cross-sectional area after the first maximum. Figure 1 shows the local minimum at distance 25 for the full dataset; if used as a fixed global cutoff, it is fitted to the dataset. Section 2.1.
  • moving average window for aortic root detection = not specified
    The paper states a moving average is used to smooth the area curve but gives no window size, which is a free choice affecting the extracted root label. Section 2.1.
  • gamma in focal and focal skeleton recall losses = 2
    Set as in Lin et al. [10]. This hyperparameter controls the strength of the focal weighting and directly affects the proposed loss. Section 3, Implementation details.
  • Dice and cross-entropy loss weights = 0.25 and 0.75
    Used in the DiceCE baseline and in combination with focal loss; chosen without reported tuning. Section 3.
assumptions (4)
  • domain assumption TotalSegmentator coarse labels for aorta, left ventricle, and iliac arteries are accurate enough to serve as bases for deriving TAVR-relevant structures.
    The pseudo-label pipeline applies distance transforms and area analysis on top of the TotalSegmentator segmentations, assuming these source labels are correct. Section 2.1.
  • domain assumption The aortic root corresponds to the portion of the aorta between the annulus and the first local minimum of cross-sectional area after the first maximum.
    This anatomical heuristic is used to cut the aortic root label and is not validated against expert definitions. Section 2.1 and Figure 1.
  • domain assumption Pseudo-labels generated by these rules are a valid proxy for true anatomy when training and evaluating segmentation models.
    No expert manual annotations are used to confirm that the derived valve, annulus, and aortic root labels match clinical measurements. Section 2.1 and Section 4.
  • domain assumption Excluding TotalSegmentator CTs that 'do not show TAVR-relevant anatomy' does not bias the benchmark.
    The paper omits 650 of 1228 cases without stating objective inclusion criteria, and the effect of this selection on anatomy variety is not analyzed. Section 2.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic Segmentation for Preoperative Planning in Transcatheter Aortic Valve Replacement." pith.science (2026). https://pith.science/paper/2LMX3IIO

@misc{pith2026250716573,
  author       = {Pith},
  title        = {Pith review of: Semantic Segmentation for Preoperative Planning in Transcatheter Aortic Valve Replacement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LMX3IIO}},
  note         = {Machine review of arXiv:2507.16573}
}
read the original abstract

When preoperative planning for surgeries is conducted on the basis of medical images, artificial intelligence methods can support medical doctors during assessment. In this work, we consider medical guidelines for preoperative planning of the transcatheter aortic valve replacement (TAVR) and identify tasks, that may be supported via semantic segmentation models by making relevant anatomical structures measurable in computed tomography scans. We first derive fine-grained TAVR-relevant pseudo-labels from coarse-grained anatomical information, in order to train segmentation models and quantify how well they are able to find these structures in the scans. Furthermore, we propose an adaptation to the loss function in training these segmentation models and through this achieve a +1.27% Dice increase in performance. Our fine-grained TAVR-relevant pseudo-labels and the computed tomography scans we build upon are available at https://doi.org/10.5281/zenodo.16274176.

Figures

Figures reproduced from arXiv: 2507.16573 by the authors.

Figure 1
Figure 1. The number of voxels in the cross-section of the aorta over the entire [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Comparison between target and predictions of Unet, Vnet and Swin UN [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 11 canonical work pages

  1. [1]

    EHJ (2022)

    Vahanian et al., A.: 2021 esc/eacts guidelines for the management of valvular heart disease. EHJ (2022)

    work page 2022

  2. [2]

    in ICIP (2024)

    Jaus et al., A.: Towards Unifying Anatomy Segmentation. in ICIP (2024)

    work page 2024

  3. [3]

    In: MICCAI Workshop (2022)

    Hatamizadeh et al., A.: Swin UNETR: Swin Transformers for Semantic Seg- mentation of Brain Tumors in MRI Images. In: MICCAI Workshop (2022). https://doi.org/10.48550/arXiv.2201.01266, http://arxiv.org/abs/2201.01266

    arXiv doi:10.48550/arxiv.2201.01266 2022

  4. [4]

    3DIMPVT (2016)

    Milletari et al., F.: V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. 3DIMPVT (2016). https://doi.org/10.48550/arXiv.1606.04797, http://arxiv.org/abs/1606.04797

    arXiv doi:10.48550/arxiv.1606.04797 2016

  5. [5]

    Radiology: Ar- tificial Intelligence (2023)

    Wasserthal et al., J.: TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Ar- tificial Intelligence (2023). https://doi.org/10.1148/ryai.230024, https://pubs.rsna.org/doi/10.1148/ryai.230024

    doi:10.1148/ryai.230024 2023

  6. [6]

    Cardoso et al., M.J.: Monai: An open-source framework for deep learning in health- care (2022)

    work page 2022

  7. [7]

    Circulation: Cardiovascular Inter- ventions (2008) 10 C

    Piazza et al., N.: Anatomy of the aortic valvar complex and its ifmplications for transcatheter implantation of the aortic valve. Circulation: Cardiovascular Inter- ventions (2008) 10 C. Zöllner et al

    work page 2008

  8. [8]

    IEEE TMI (2017)

    Lessmann et al., N.: Automatic calcium scoring in low-dose chest ct using deep neural networks with dilated convolutions. IEEE TMI (2017)

    work page 2017

Show all 17 references
  1. [9]

    MICCAI (2015)

    Ronneberger et al., O.: U-Net: Convolutional Networks for Biomedical Image Seg- mentation. MICCAI (2015). https://doi.org/10.1007/978-3-319-24574-4_28

    2015 doi

  2. [10]

    In: CVPR (2018), http://arxiv.org/abs/1708.02002

    Lin et al., T.Y.: Focal Loss for Dense Object Detection. In: CVPR (2018), http://arxiv.org/abs/1708.02002

    2018 arXiv

  3. [11]

    The lancet (2006)

    Nkomo et al., V.T.: Burden of valvular heart diseases: a population-based study. The lancet (2006)

    2006

  4. [12]

    Medi- cal Image Analysis (2019)

    Zhuang et al., X.: Evaluation of algorithms for Multi-Modality Whole Heart Segmentation: An open-access grand challenge. Medi- cal Image Analysis (2019). https://doi.org/10.1016/j.media.2019.101537, https://www.sciencedirect.com/science/article/pii/S1361841519300751

    2019

  5. [13]

    In: ECCV (2024), http://arxiv.org/abs/2404.03010

    Kirchhoff et al., Y.: Skeleton Recall Loss for Connectivity Conserving and Re- source Efficient Segmentation of Thin Tubular Structures. In: ECCV (2024), http://arxiv.org/abs/2404.03010

    2024 arXiv

  6. [14]

    In: International Workshop on Statistical Atlases and Computational Models of the Heart

    Almar-Munoz, E., Pamminger, M., Kremser, C., Haltmeier, M., Mayr, A.: Be- yond the standards: Fully-automated aortic annulus segmentation on contrast-free magnetic resonance imaging using a computational aorta unwrapping method. In: International Workshop on Statistical Atlase...

    2024

  7. [15]

    In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA)

    Lambert, Z., Petitjean, C., Dubray, B., Kuan, S.: Segthor: Segmentation of thoracic organs at risk in ct images. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA). pp. 1–6. IEEE (2020)

    2020

  8. [16]

    Pak, D.H., Caballero, A., Sun, W., Duncan, J.S.: Efficient aortic valve multilabel segmentationusingaspatialtransformernetwork.In:2020IEEE17thinternational symposium on biomedical imaging (ISBI). pp. 1738–1742. IEEE (2020)

    2020

  9. [17]

    In:MedicalImageComputingandComputer-AssistedIntervention,Beijing,China, September 20-24, 2010, Proceedings, Part I 13

    Zheng, Y., John, M., Liao, R., Boese, J., Kirschstein, U., Georgescu, B., Zhou, S.K., Kempfert, J., Walther, T., Brockmann, G., et al.: Automatic aorta segmentation and valve landmark detection in c-arm ct: application to aortic valve implantation. In:MedicalImageComputingandC...

    2010

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.