Pith. sign in

REVIEW 4 major objections 8 minor 47 references

Shape-Based Inductive Bias for Glioma Grading from Tumor Contours

T0 review · 4 major / 8 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Aligned Fourier shape tokens let a tiny MLP beat pixel CNNs and ViTs at glioma grading from tumor contours, with far fewer parameters.

desk verdict Careful shape-token recipe with real parameter efficiency on BraTS contours; the “MLP wins” headline is softer than the efficiency story. read the letter →

arxiv 2607.26090 v1 pith:K3RZL3T7 submitted 2026-07-27 eess.IV cs.CVcs.LG

Puneet Velidi , Michelle F. Miranda , Farouk Nathoo , Ashery Mbilinyi , Cédric Beaulac This is my paper
classification eess.IVcs.CVcs.LG
keywords gliomagradingshapetokensFourierdescriptorsfunctionalalignmentinductivebiasBraTScontourclassificationdimensionalityreduction
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

When the scientific signal is tumor shape rather than image texture, rasterizing masks and training large vision models wastes capacity on nuisances that classical shape analysis already knows how to remove. This paper aligns closed tumor contours, splits global pose from residual Fourier harmonics, and packs the result into a few hundred frequency-ordered tokens. On BraTS 2020 under patient-disjoint cross-validation, a compact MLP on those tokens reaches the best mean balanced accuracy and the best low-grade F1, while using tens of thousands of parameters instead of millions. A controlled simulation where labels are built from deformation and mid-frequency shape confirms the same gap: shape models climb well above chance while pixel models stay near it when pose is stripped from the raster. The practical payoff is a fixed-size, geometrically meaningful representation that is cheaper to store and train and easier to interpret by harmonic band.

What carries the argument

Shape tokens: after functional alignment of a closed contour, one six-value deformation token (translation, inverse scale, phase, rotation) plus one six-value residual Fourier token per harmonic (signed coefficients and amplitudes), frequency-ordered into a matrix of at most 198 scalars that an MLP flattens or a small transformer attends over.

What would settle it

Re-run the same patient-disjoint protocol with true histopathologic or molecular grade labels, or with multi-slice / surface contours instead of the single largest axial slice; if the MLP’s balanced-accuracy and LGG-F1 edges over the pixel baselines disappear, the central claim fails.

Watch

Extended reading notes

Core claim

On BraTS 2020 tumor contours under five-fold patient-disjoint cross-validation with grouped inner model selection, a compact MLP on at most 198 aligned Fourier shape-token scalars attains the highest mean balanced accuracy (71.5%) and mean LGG F1 (54.9%), beating ResNet-18 and ViT-Tiny while using at least 46 times fewer parameters; pooled out-of-fold balanced accuracy is 72.4% (patient-bootstrap 95% CI 66.4–77.8%).

Load-bearing premise

The grade label is an imaging phenotype defined by whether a 2D connected component contains the full triad of necrosis, edema, and enhancing tumor—not histopathologic or molecular grade—and each 3D volume is reduced to its largest-area axial contour.

Editorial extensions

If this is right

  • Morphology-driven medical tasks can start from a few hundred aligned Fourier scalars rather than full-resolution rasters, cutting parameters and training time by large factors.
  • Deformation parameters and harmonic bands become directly ablatable features, supporting mechanistic checks of which scales of shape drive the decision.
  • Storage and CPU inference cost become independent of original image resolution once masks are turned into fixed-size token matrices.
  • When labels truly depend on pose plus mid-frequency boundary detail, pixel models that recenter and rescale can lose the signal that shape tokens keep.
  • The same pipeline can be paired with automatic segmenters to grade morphology without hand-drawn masks.

Reading between the lines

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

  • Coastlines, field boundaries, fire perimeters, and urban footprints are natural next domains: the same tokens would turn massive geospatial rasters into compact, auditable geometric summaries.
  • If higher harmonics prove unstable across scanners while low-order deformation and elongation stay predictive, the method supplies a built-in frequency filter for multi-site studies.
  • Attention maps that shift under contour noise already hint that transformers may be most useful precisely when label-relevant interactions cross harmonic scales—an experiment the paper leaves open.
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 / 8 minor

Summary. The manuscript proposes representing tumor contours as "shape tokens": a six-component deformation token (translation, inverse scale, phase, rotation) plus one six-component residual token per Fourier harmonic, obtained via the functional shape-alignment framework of Moindjié et al. On BraTS 2020, each 3D segmentation is reduced to the largest-area axial slice, connected components are converted to arc-length-resampled closed contours, and a binary HGG/LGG-like target is derived from co-occurrence of mask subregion labels 1, 2, and 4. Under five-fold patient-disjoint cross-validation with grouped inner-split model selection, a compact MLP over at most 198 scalars achieves mean balanced accuracy 71.5% (vs. 65.9% ResNet-18, 63.3% ViT-Tiny) and mean LGG F1 54.9%, with 2.9k–117.3k parameters; pooled out-of-fold balanced accuracy is 72.4% (patient-bootstrap 95% CI 66.4–77.8%). A controlled simulation in which labels are built from deformation and mid-frequency harmonic terms shows shape-based models above chance while pixel baselines on recentered/rescaled rasters sit at chance. The authors claim dimensionality reduction, competitive accuracy, and improved minority-class detection from a shape-based inductive bias.

Significance. If the comparative results hold, the paper makes a useful and credible point: for small labeled cohorts where the signal is morphological, an explicitly aligned Fourier representation delivers class-balanced performance competitive with much larger pixel models at a ≥46× parameter reduction and orders-of-magnitude lower compute, with interpretable harmonic axes. The evaluation hygiene is a genuine strength relative to typical papers at this venue: patient-grouped outer folds with grouped inner selection, normalization and reference refit on training data only, patient-level bootstrap intervals on pooled OOF predictions, and an implementation-path equivalence check (single vs. double Fourier path, Supplement C). The controlled simulation, while deliberately constructed to favor the token representation, is a transparent diagnostic of what information each representation preserves. The contribution is methodological and incremental rather than clinical, and the claims are correspondingly strongest in their dimensionality-reduction/efficiency dimension and weakest in their head-to-head accuracy dimension.

major comments (4)
  1. [Table 3 / §6.1] Table 3 and §6.1: the headline comparison is not shown to exceed fold/selection variability. The MLP mean balanced accuracy is 71.5 (SD 6.4) versus ResNet-18 65.9 (SD 4.6); a simple unpaired calculation on five folds gives a difference SE around 3.5–3.7 (t ≈ 1.6), and the same concern is sharper for Table 4 (LGG F1 54.9 (SD 12.2) vs 48.4 (SD 11.4), with roughly 19 LGG contours per outer test fold). Because all models share identical outer folds and every contour has exactly one OOF prediction, a paired patient-level bootstrap of the OOF metric difference (MLP minus each baseline) is directly computable from artifacts the authors already produce for the pooled CI (§3.1). The paper should either report such a paired interval/test for the differences in balanced accuracy and LGG F1, or soften the abstract/§6.1 wording from "highest"/"superior" to "competitive." As written, the abstract's le
  2. [§3.1 / Tasks] §3.1 (target definition): the binary label is constructed from co-occurrence of mask subregion labels 1, 2, and 4 within a component, not from the BraTS-provided patient grade. The manuscript is transparent that this is an "operational imaging-phenotype classification," but the title and abstract claim "glioma grading," and the proxy has a structural relationship with morphology: components large or heterogeneous enough to contain all three subregions will systematically differ in contour scale and complexity from those that do not, so part of what the shape model detects may be the label-construction rule rather than grade-related morphology. Since BraTS 2020 ships actual HGG/LGG patient labels, the load-bearing fix is cheap: report the concordance between the proxy and the true grade labels, and ideally report the same Table 3/Table 4 pipeline on the true labels (or explain why the pro
  3. [§4.1] §4.1 (reference template): the alignment reference μ is "the first preshape in the deterministic inner-training order" (recomputed on the outer-training fold after selection), an arbitrary, order-dependent choice, and the manuscript states that "sensitivity to the reference choice remains to be tested." Since every downstream token is a residual against μ, fragility here would undermine the entire representation. A minimal sensitivity check—e.g., re-running one or two outer folds with several random training-set references, or with a training-fold Fréchet-type mean from [28]—would suffice to bound this; the authors already flag it as future work, but given that alignment is the paper's core operation, at least a small empirical bound belongs in this revision.
  4. [§3.2 / §6.2 / Table 9] §3.2, Eq. (2), and Table 2 vs. the abstract: in the simulation the label depends on log s and sign(t_x), which the raster pipeline explicitly removes ("recentering and rescaling"), while the deformation token retains them. The text is commendably honest about this ("the unequal information retained by the token and raster representations limits broader architecture-level conclusions"), but the abstract sentence "shape-based models reach 56.3–71.5% balanced accuracy while the pixel models remain at 50.0–52.5%" presents the contrast without the caveat and invites an architecture-level reading the design cannot support. Either the abstract should carry the information-asymmetry caveat, or the paper should add a matched-information control (rasters rendered without recentering/rescaling, or with scale/translation given to the pixel models as side inputs). Relatedly, the rise of the CNN from
minor comments (8)
  1. [various] Typos/grammar: "tokenizaiton" (§4, first line); "su press" (§2.1); "allow the classifier learn" (§2, end); "in of itself" (§6.2); "Simulation pixel-based baselines uses" (§4.3).
  2. [Table 5 / Table 6] Table 5 reports mean final-training wall-clock time of 0.16 s for the MLP, while Table 6 reports 0.75 s for a fixed-split MLP run; please clarify what Table 5 measures (per-epoch? selected-epoch totals?) since a sub-second total training time is surprising and the efficiency claim ("at least 2,000 times less training time") depends on it.
  3. [§3.2 / Table 2] Table 2 and Table 9 confidence intervals are t-intervals over 20 training initializations with a fixed 2,000-sample test set; they therefore quantify initialization variance only, not variability over data generation. A sentence making this explicit would prevent over-reading of the simulation CIs.
  4. [§4.2] §4.2, Eq. (7): the zero padding in the deformation token and its role as position-0 readout could use one more sentence of rationale (why a constant pad rather than, e.g., a second derived quantity), particularly for the Shape Transformer where the token dimension feeds the projection.
  5. [Table 1] Table 1 sorts selected K values "for compactness," which destroys the fold-to-configuration correspondence; a per-fold listing (fold index × selected K, width, dropout) would be more informative given the fold-adaptivity discussion in §5.
  6. [§3.2] Eq. (1): the indicator term multiplying the harmonic sum is typeset inline and is easy to misread as part of the summand; please reformat.
  7. [Figures 1 and 4] Figure 1: contour examples are illustrative but small; axis scale bars or a shared size reference would help readers see the size difference the text emphasizes. Similarly, Figure 4's double-panel overlay with interpolation makes the per-harmonic values hard to read; consider discrete markers at integer harmonics.
  8. [Code availability] The promised anonymized repository should include the outer/inner fold assignments and the stored OOF predictions, since the paired comparison requested above depends on them; please confirm in the revision.

Circularity Check

1 steps flagged · score 2.0 of 10

BraTS results are not circular; only the controlled simulation builds labels from the same deformation/harmonic cues the tokenizer exposes (disclosed diagnostic, not a forced empirical claim).

  1. self definitional [Section 3.2, Eqs. (1)–(2) and following paragraph]
    "Labels deliberately combine information from the deformation and latent-shape components, η=log s + 0.5 sign(t_x) + 1.2(cls−0.5)+ϵ, y=1{η>0}. The terms log s and sign(t_x) are available to a shape-based classifier through the deformation token, while cls is encoded by the mid-frequency residual shape tokens. Pixel-based classifiers read a 64×64 filled raster produced after recentering and rescaling the polygon"

    In the noise-free simulation, the label is defined directly from deformation scalars and the latent mid-frequency class indicator that the shape tokenizer is built to retain, while standard raster preprocessing removes those deformation variables. Above-chance shape accuracy and near-chance pixel accuracy are therefore partly guaranteed by how y and the two input pipelines are constructed. The paper discloses this as a controlled diagnostic; it does not underwrite the BraTS claim by the same reduction.

full rationale

This is an empirical representation-comparison paper, not a first-principles derivation. On BraTS, the binary target is defined from mask subregion co-occurrence (full triad of labels 1/2/4 → HGG; otherwise LGG), inputs are aligned Fourier tokens from contours, and evaluation is patient-disjoint outer CV with grouped inner selection—none of which makes accuracy equal the inputs by construction. The functional alignment is taken from Moindjié, Beaulac & Descary [28] (overlapping coauthor Beaulac); that is ordinary methodological reuse, not a uniqueness theorem that forbids alternatives or forces the reported metrics. The only by-construction element is the hybrid simulation: y is an explicit function of log-scale, sign(t_x), and a latent harmonic class that residual tokens carry, while raster baselines are recentered/rescaled so those deformation cues are stripped. The paper states this design openly and treats the run as a diagnostic of information retention, not as independent external proof of architecture superiority. That warrants a low score (≈2), not a claim that the central BraTS finding is circular.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The load-bearing content is mostly standard Fourier/Procrustes alignment plus an operational BraTS phenotype rule and many selection hyperparameters. Independent geometric content comes from prior functional-shape work; the paper adds the token interface, classifiers, and empirical protocol rather than new physical entities.

free parameters (8)
  • Fourier bandwidth K = selected per fold in {8,16,24,32}
    Controls input dimension and which residual harmonics enter the classifier; chosen by inner validation, not fixed a priori.
  • MLP width and dropout = dh in {32..512} sweep; selected 32–256, dropout in {0,0.1,0.2}
    Capacity and regularization fitted on inner validation; selected configs span 2.9k–117.3k parameters.
  • Shape Transformer depth/width/heads = L in {4,6}, d in {128,192,256}, H in {4,8}
    Architecture hyperparameters selected per outer fold on inner validation.
  • Contour resample size P = 256
    Hand-chosen uniform arc-length grid before Fourier projection; affects aliasing and coefficient estimates.
  • Phase grid resolution = 101-point δ grid
    Discretizes starting-point alignment; practical accuracy/cost tradeoff stated by authors.
  • Reference template μ = first training preshape (deterministic order)
    Alignment target chosen ad hoc rather than Fréchet mean; authors flag sensitivity as future work.
  • Simulation label coefficients = η = log s + 0.5 sign(tx) + 1.2(cls−0.5) + ε, ε~N(0,0.1)
    Hand-set weights define how much deformation vs mid-frequency shape drives y; determines the diagnostic task.
  • Minimum component size and slice rule = ≥50 pixels; largest total tumor area axial slice
    Defines which 2D contours enter the dataset and thus the empirical claim.
assumptions (5)
  • domain assumption Closed planar contours admit the deformation model C(t)=ρ Oθ C̃∘γδ(t)+T with Fourier-block phase/rotation alignment as in Moindjié et al.
    Technical basis stated in Introduction and §4.1; residual tokens and deformation token inherit meaning from this framework.
  • ad hoc to paper Co-occurrence of mask labels 1,2,4 defines HGG-like phenotype; other combinations define LGG-like phenotype for this study.
    Tasks §3.1 explicitly constructs an operational target rather than histopathology; central BraTS claim uses this rule.
  • domain assumption Largest-area axial 2D connected-component contours carry sufficient morphological signal for the phenotype task.
    From-3D-to-2D pipeline in §3.1; multi-slice/surface extension left to future work.
  • domain assumption Expert BraTS segmentations are accurate enough that contour geometry is not dominated by annotation noise for grading-from-shape.
    Analysis begins from given masks; SAM-based masks discussed only as future applicability.
  • standard math Standard supervised learning metrics under patient-grouped CV estimate out-of-sample phenotype discrimination.
    Usual i.i.d.-within-patient-block assumption for medical CV; bootstrap resamples patients with their contours.
invented entities (1)
  • Shape tokens (6-D deformation token + per-harmonic residual coefficient/amplitude tokens)
    purpose: Provide a fixed-size, frequency-ordered input for MLP and Transformer classifiers after functional alignment.
    Representation interface introduced in Contributions and §4.2; geometric meaning of harmonics is classical, but this token matrix and deformation-as-position-0 readout are paper-specific packaging without external falsifiable prediction beyond the reported tables.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Shape-Based Inductive Bias for Glioma Grading from Tumor Contours." pith.science (2026). https://pith.science/paper/K3RZL3T7

@misc{pith2026260726090,
  author       = {Pith},
  title        = {Pith review of: Shape-Based Inductive Bias for Glioma Grading from Tumor Contours},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K3RZL3T7}},
  note         = {Machine review of arXiv:2607.26090}
}
read the original abstract

Glioma grading from tumor contours is often treated as a pixel problem even when the signal of interest is shape. We align closed contours with a functional shape-alignment framework, separate global deformation from residual Fourier shape, and organize these quantities as frequency-ordered tokens. In five-fold patient-disjoint cross-validation on BraTS~2020 tumor contours, with model selection performed using grouped inner validation, a compact multilayer perceptron (MLP) achieves the highest mean balanced accuracy at 71.5\%, compared with 65.9\% for ResNet-18 and 63.3\% for ViT-Tiny. It also gives the highest mean low-grade glioma F1 at 54.9\%. Its pooled out-of-fold balanced accuracy is 72.4\% (patient-bootstrap 95\% CI: 66.4--77.8\%). The selected MLPs use 2.9k--117.3k parameters across folds, at least 46 times fewer than the pixel baselines. In a controlled noise-free simulation, shape-based models reach 56.3--71.5\% balanced accuracy while the pixel models remain at 50.0--52.5\%. This work demonstrates how incorporating a shape-based inductive bias at the representation level can improve interpretability and scalability while enabling substantial dimensionality reduction.

Figures

Figures reproduced from arXiv: 2607.26090 by the authors.

Figure 1
Figure 1. Tumor contours extracted from BraTS 2020 segmen [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Contour-token pipeline. The contour is converted to a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Effect of varying a single Fourier coefficient [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Paired noisy-minus-clean change in pose-query attention to harmonics, in attention percentage points, for the single and double [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Example simulated shapes. Each cell shows a [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 3 linked inside Pith

  1. [28]

    A functional approach to curve alignment and shape analysis.arXiv preprint, 2025

    Issam-Ali Moindji ´e, C ´edric Beaulac, and Marie-H ´el`ene Descary. A functional approach to curve alignment and shape analysis.arXiv preprint, 2025. 1, 2, 3, 4, 5, 8, 11

    2025

  2. [1]

    Hugo J. W. L. Aerts, Emmanuel Rios Velazquez, Ralph T. H. Leijenaar, Chintan Parmar, Patrick Grossmann, Sara Car- valho, Johan Bussink, Ren ´e Monshouwer, Benjamin Haibe- Kains, Dani ¨elle Rietveld, Frank Hoebers, Michelle M. Ri- etbergen, C. Ren ´e Leemans, Andre Dekker, John Quacken- bush, Robert J. Gillies, and Philippe Lambin. Decoding tu- mour phenot...

    2014

  3. [2]

    Kirby, John B

    Spyridon Bakas, Hamed Akbari, Aristeidis Sotiras, Michel Bilello, Martin Rozycki, Justin S. Kirby, John B. Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing the Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features.Scientific Data, 4(1):170117, 2017. 1, 3

    2017

  4. [3]

    Shape con- text: A new descriptor for shape matching and object recog- nition

    Serge Belongie, Jitendra Malik, and Jan Puzicha. Shape con- text: A new descriptor for shape matching and object recog- nition. InNeurIPS, pages 831–837, 2000. 2

    2000

  5. [4]

    Shape matching and object recognition using shape contexts.IEEE TPAMI, 24(4):509–522, 2002

    Serge Belongie, Jitendra Malik, and Jan Puzicha. Shape matching and object recognition using shape contexts.IEEE TPAMI, 24(4):509–522, 2002. 2

    2002

  6. [5]

    Valerio Biscione and Jeffrey S. Bowers. Convolutional neu- ral networks are not invariant to translation, but they can learn to be.ArXiv, abs/2110.05861, 2021. 3

    arXiv 2021

  7. [6]

    Bookstein.Morphometric Tools for Landmark Data: Geometry and Biology

    Fred L. Bookstein.Morphometric Tools for Landmark Data: Geometry and Biology. Cambridge University Press, 1991. 2

    1991

  8. [7]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.arXiv preprint, 2021. 1

    2021

Show all 47 references
  1. [8]

    Sam 3: Segment anything with concepts, 2026

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoub- hik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R¨adle, Trian...

    2026

  2. [9]

    Classification of the glioma grading using radiomics analysis.PeerJ, 6:e5982, 2018

    Hwan-Ho Cho, Seung-Hak Lee, Jonghoon Kim, and Hyunjin Park. Classification of the glioma grading using radiomics analysis.PeerJ, 6:e5982, 2018. 2

    2018

  3. [10]

    Cohen and Max Welling

    Taco S. Cohen and Max Welling. Group equivariant convo- lutional networks. InInternational Conference on Machine Learning, pages 2990–2999, 2016. 1

    2016

  4. [11]

    Cootes, Christopher J

    Timothy F. Cootes, Christopher J. Taylor, David H. Cooper, and Jim Graham. Active shape models—their training and application.Computer Vision and Image Understanding, 61 (1):38–59, 1995. 2

    1995

  5. [12]

    BERT: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Tuptanova. BERT: Pre-training of deep bidirectional trans- formers for language understanding. InNAACL-HLT, pages 4171–4186, 2019. 3

    2019

  6. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  7. [14]

    Dryden and Kanti V

    Ian L. Dryden and Kanti V . Mardia.Statistical Shape Analy- sis. John Wiley & Sons, Chichester, 1998. 1, 2

    1998

  8. [15]

    On the encoding of arbitrary geomet- ric configurations.IRE Trans

    Herbert Freeman. On the encoding of arbitrary geomet- ric configurations.IRE Trans. Electronic Computers, 10(2): 260–268, 1961. 2

    1961

  9. [16]

    Imagenet-trained cnns are biased towards texture; increas- ing shape bias improves accuracy and robustness.ArXiv, abs/1811.12231, 2018

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increas- ing shape bias improves accuracy and robustness.ArXiv, abs/1811.12231, 2018. 3

    2018 arXiv

  10. [17]

    Gillies, Paul E

    Robert J. Gillies, Paul E. Kinahan, and Hedvig Hricak. Ra- diomics: Images are more than pictures, they are data.Radi- ology, 278(2):563–577, 2016. 1, 2

    2016

  11. [18]

    Roth, and Daguang Xu

    Ali Hatamizadeh, Dong Yang, Holger R. Roth, and Daguang Xu. UNETR: Transformers for 3D medical image segmen- tation. InIEEE/CVF Winter Conference on Applications of Computer Vision, pages 1748–1758, 2022. 2

    2022

  12. [19]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR, pages 770–778, 2016. 6

    2016

  13. [20]

    Jaeger, Simon A

    Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Klaus H. Maier-Hein. nnU-Net: a self- configuring method for deep learning-based biomedical im- age segmentation.Nature Methods, 18:203–211, 2021. 2

    2021

  14. [21]

    Kuhl and Charles R

    Frank P. Kuhl and Charles R. Giardina. Elliptic Fourier fea- tures of a closed contour.Computer Graphics and Image Processing, 18(3):236–258, 1982. 1, 2

    1982

  15. [22]

    Philippe Lambin, Emmanuel Rios-Velazquez, Ralph Leije- naar, Sara Carvalho, Ruud G. P. M. van Stiphout, Patrick Granton, Catharina M. L. Zegers, Robert Gillies, Ronald Boellard, Andre Dekker, and Hugo J. W. L. Aerts. Ra- diomics: Extracting more information from medical images...

    2012

  16. [23]

    Kosiorek, Seungjin Choi, and Yee Whye Teh

    Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. InInternational Conference on Machine Learn- ing, pages 3744–3753, 2019. 3

    2019

  17. [24]

    Geospatial big data handling theory and methods: A review and research challenges.ISPRS Journal of Pho- togrammetry and Remote Sensing, 115:119–133, 2016

    Songnian Li, Suzana Dragicevic, Francesc Ant ´on Castro, Monika Sester, Stephan Winter, Arzu Coltekin, Christopher Pettit, Bin Jiang, James Haworth, Alfred Stein, and Tao Cheng. Geospatial big data handling theory and methods: A review and research challenges.ISPRS Journal of ...

    2016

  18. [25]

    Lorensen and Harvey E

    William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In ACM SIGGRAPH, pages 163–169, 1987. 3

    1987

  19. [26]

    Segment anything in medical images

    Jun Ma and Bo Wang. Segment anything in medical images. ArXiv, abs/2304.12306, 2023. 8

    2023 arXiv

  20. [27]

    Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al

    Bjoern H. Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal Brain Tumor image segmentation benchmark (BRATS).IEEE Trans. Medical Imaging, 34 (10):1993–...

    1993

  21. [29]

    Multivariate planar curves: A statistical framework for shape analysis in images.arXiv preprint, 2025

    Issam-Ali Moindji ´e, Marie-H ´el`ene Descary, and C ´edric Beaulac. Multivariate planar curves: A statistical framework for shape analysis in images.arXiv preprint, 2025. 2

    2025

  22. [30]

    Shape discrimination using fourier descriptors.IEEE Trans

    Eric Persoon and King-Sun Fu. Shape discrimination using fourier descriptors.IEEE Trans. Systems, Man, and Cyber- netics, 7(3):170–179, 1977. 2

    1977

  23. [31]

    James Rohlf and Dennis Slice

    F. James Rohlf and Dennis Slice. Extensions of the Pro- crustes method for the optimal superimposition of land- marks.Systematic Zoology, 39(1):40–59, 1990. 2

    1990

  24. [32]

    U- Net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional networks for biomedical image segmen- tation. InMedical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234–241. Springer,

    2015

  25. [33]

    Maurer, Michael Syn- owitz, Joost W¨ustefeld, Tim Marnitz, Florian Streitparth, and Edzard Wiener

    Max-Ludwig Sch ¨afer, Martin H. Maurer, Michael Syn- owitz, Joost W¨ustefeld, Tim Marnitz, Florian Streitparth, and Edzard Wiener. Low-grade (WHO II) and anaplastic (WHO III) gliomas: differences in morphology and MRI signal in- tensities.European Radiology, 23(10):2846–2853, ...

    2013

  26. [34]

    skimage.measure.find contours

    scikit-image developers. skimage.measure.find contours. scikit-image API documentation, 2026. Accessed 2026-07-

    2026

  27. [35]

    Klassen.Functional and Shape Data Analysis

    Anuj Srivastava and Eric P. Klassen.Functional and Shape Data Analysis. Springer, 2016. 1, 2

    2016

  28. [36]

    Joshi, and Ian H

    Anuj Srivastava, Eric Klassen, Shantanu H. Joshi, and Ian H. Jermyn. Shape analysis of elastic curves in euclidean spaces. IEEE TPAMI, 33(7):1415–1428, 2011. 1, 2

    2011

  29. [37]

    Deep learn- ing with functional inputs.Journal of Computational and Graphical Statistics, 32(1):171–180, 2023

    Barinder Thind, Kevin Multani, and Jiguo Cao. Deep learn- ing with functional inputs.Journal of Computational and Graphical Statistics, 32(1):171–180, 2023. 2

    2023

  30. [38]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InNeurIPS, pages 5998–6008, 2017. 3, 5

    2017

  31. [39]

    Neural networks for scalar input and functional output.Statistics and Com- puting, 33(5):118, 2023

    Sidi Wu, C ´edric Beaulac, and Jiguo Cao. Neural networks for scalar input and functional output.Statistics and Com- puting, 33(5):118, 2023. 2

    2023

  32. [40]

    Functional autoen- coder for smoothing and representation learning.Statistics and Computing, 34(6):203, 2024

    Sidi Wu, C ´edric Beaulac, and Jiguo Cao. Functional autoen- coder for smoothing and representation learning.Statistics and Computing, 34(6):203, 2024. 2

    2024

  33. [41]

    An automated ap- proach for predicting glioma grade and survival of LGG pa- tients using CNN and radiomics.Frontiers in Oncology, 12: 969907, 2022

    Cheng Xu, Yue Peng, Wensheng Zhu, Zhen Chen, Jing Li, Wei Tan, Zhen Zhang, and Xiang Chen. An automated ap- proach for predicting glioma grade and survival of LGG pa- tients using CNN and radiomics.Frontiers in Oncology, 12: 969907, 2022. 2

    2022

  34. [42]

    Zacharaki, Sumei Wang, Sanjeev Chawla, Dong Soo Yoo, Ronald Wolf, Elias R

    Evangelia I. Zacharaki, Sumei Wang, Sanjeev Chawla, Dong Soo Yoo, Ronald Wolf, Elias R. Melhem, and Chris- tos Davatzikos. Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magnetic Resonance in Medicine, 62(6):1609–1618, 2009. 2

    2009

  35. [43]

    Zahn and Ralph Z

    Charles T. Zahn and Ralph Z. Roskies. Fourier descriptors for plane closed curves.IEEE Trans. Computers, 21(3):269– 281, 1972. 1, 2

    1972

  36. [44]

    Mazurowski

    Yixin Zhang and Maciej A. Mazurowski. Convolutional neu- ral networks rarely learn shape for semantic segmentation. Pattern Recognit., 146:110018, 2023. 3

    2023

  37. [45]

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer. InIEEE/CVF Interna- tional Conference on Computer Vision, pages 16259–16268,

  38. [46]

    Deep learning in remote sensing: A comprehensive review and list of resources.IEEE Geoscience and Remote Sensing Maga- zine, 5(4):8–36, 2017

    Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources.IEEE Geoscience and Remote Sensing Maga- zine, 5(4):8–36, 2017. 8

    2017

  39. [47]

    Cheng, An- dra V

    Ying Zhuge, Holly Ning, Peter Mathen, Jason Y . Cheng, An- dra V . Krauze, Kevin Camphausen, and Robert W. Miller. Automated glioma grading on conventional MRI images us- ing deep convolutional neural networks.Medical Physics, 47 (7):3044–3053, 2020. 2 Supplementary Material A...

    2020

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.