pith. sign in

arxiv: 2607.02133 · v1 · pith:DFPKV36Hnew · submitted 2026-07-02 · 📊 stat.AP

Quaternion Nondecimated Wavelet Descriptors for Multiclass Breast Histology Classification

Sara Antonijevic , Brani Vidakovic This is my paper

Pith reviewed 2026-07-03 03:02 UTC · model grok-4.3

classification 📊 stat.AP
keywords quaternion waveletsbreast histology classificationBACH datasetnondecimated wavelet transforminterpretable featurescolor image processingmulticlass tissue recognition
0
0 comments X

The pith

Encoding RGB histology images as quaternion fields and applying nondecimated wavelet transforms yields balanced four-class breast tissue classification on the BACH dataset.

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

The paper establishes a framework that encodes each RGB breast histology image as a pure quaternion field and applies a two-dimensional quaternion nondecimated wavelet transform to produce multiscale directional coefficient fields that keep color coupled as a single vector. From the resulting coefficients it extracts families of interpretable features that summarize stain balance, energy, amplitude heterogeneity, phase concentration, directional anisotropy, and scale-dependent decay, each linked to properties such as nuclear density or glandular organization. These descriptors are fed to a radial-kernel SVM and evaluated with repeated nested cross-validation on the balanced four-class BACH challenge set of normal, benign, in situ, and invasive tissue. The resulting classifier shows balanced performance across classes, with misclassifications mainly between adjacent categories and with permutation importance confirming that directional, phase, anisotropy, scale, and amplitude groups all contribute.

Core claim

Quaternion nondecimated wavelet descriptors derived from pure quaternion encodings of RGB images capture color-coupled multiscale geometry that supports balanced multiclass recognition of breast histology into normal, benign, in situ, and invasive categories, with the SVM drawing on directional, phase-concentration, and anisotropy properties rather than global color statistics alone.

What carries the argument

The quaternion nondecimated wavelet transform in two dimensions (QNDWT2D) applied to pure quaternion fields, which generates multiscale, directional, color-coupled coefficient maps on the original image grid.

If this is right

  • Errors concentrate among adjacent categories while normal and invasive classes are rarely reversed.
  • Permutation importance ranks directional, phase-concentration, anisotropy, scale, and amplitude-variability groups as contributory.
  • The method supplies a reproducible baseline that uses no pretrained networks or external databases.
  • Each feature family is explicitly tied to a histopathological property such as stain balance or glandular organization.

Where Pith is reading between the lines

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

  • The same quaternion-wavelet pipeline could be tested on other chromatic medical images where color coupling matters, such as dermatology or ophthalmology slides.
  • The explicit link between feature groups and tissue properties offers a route to model explanations that align with pathologist language.
  • Extending the descriptors to three-dimensional volumes or time-series histology would test whether the scale-decay and anisotropy components generalize beyond two-dimensional sections.

Load-bearing premise

Encoding RGB images as pure quaternion fields, applying QNDWT2D, and extracting the listed feature families preserves the diagnostic information in color, texture, orientation, and tissue architecture without material loss or artifact introduction.

What would settle it

A drop in balanced accuracy or a shift in error patterns when the quaternion phase-concentration or directional-anisotropy feature groups are ablated, or when the same descriptors are applied to an independent histology collection with different staining protocols.

Figures

Figures reproduced from arXiv: 2607.02133 by Brani Vidakovic, Sara Antonijevic.

Figure 1
Figure 1. Figure 1: Representative H&E histology images from the four BACH diagnostic categories: normal, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic organization of the quaternion two-dimensional nondecimated wavelet trans [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quaternion components in the diagonal hierarchy of the nondecimated wavelet transform. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Radial SVM confusion display. Rows are predicted classes and columns are true classes. Each cell shows the repeated cross-validation count and the corresponding percentage conditional on the true class. 5.3 Marginal binary summaries from the four-class classifier Although the primary analysis is a four-class problem, several marginal binary contrasts derived from the aggregated four-class confusion table a… view at source ↗
Figure 5
Figure 5. Figure 5: Group permutation importance for feature classes with mean decrease in held-out balanced accuracy greater than 5 percentage points. Importance is computed by block-permuting all features in a group within the held-out folds. The groups are overlapping rather than disjoint, since a feature can belong to both an orientation group and a level group, so the bars are marginal drops in accuracy when each view is… view at source ↗
Figure 6
Figure 6. Figure 6: Individual permutation importance for nonredundant predictors with mean decrease in held-out balanced accuracy greater than 0.3 percentage points. Predictors that were perfectly correlated with stronger predictors were removed from the display. Individual effects are smaller than group effects because the radial SVM uses many correlated predictors jointly. 6 Conclusions This paper proposed a quaternion non… view at source ↗
Figure 7
Figure 7. Figure 7: Correlation matrix for the nonredundant individual predictors whose mean permutation [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
read the original abstract

Breast histology images carry diagnostic information in color, texture, orientation, and tissue architecture across a range of scales. In H&E microscopy this information is inherently chromatic and is not fully recovered when the red, green, and blue (RGB) channels are reduced to grayscale or transformed as independent scalar images. We propose an interpretable quaternion nondecimated wavelet framework for breast histology classification. Each RGB image is encoded as a pure quaternion field, and a quaternion nondecimated wavelet transform in two dimensions (QNDWT2D) produces multiscale, directional, color-coupled coefficient fields on the original image grid, keeping color as a single vector quantity rather than three separate channels. From these coefficients we build interpretable feature families summarizing stain balance, wavelet energy, amplitude heterogeneity, quaternion phase concentration, color-axis geometry, directional anisotropy, orientation entropy, and scale-dependent energy decay, each tied to a histopathological property such as nuclear density or glandular organization. We evaluate the descriptors on the BreAst Cancer Histology (BACH) challenge, a balanced four-class set of normal, benign, in situ, and invasive tissue, using a radial-kernel support vector machine (SVM) with repeated nested cross-validation. The descriptors yield balanced recognition across classes, with errors concentrated among adjacent categories while normal and invasive are rarely reversed. Permutation importance shows that directional, phase-concentration, anisotropy, scale, and amplitude-variability groups all contribute, indicating that the classifier draws on genuine quaternion and multiscale geometry rather than global color alone. The framework uses no pretrained networks, learned filters, or external databases, offering a reproducible, interpretable baseline for computational pathology.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes encoding RGB breast histology images as pure quaternion fields, applying a quaternion nondecimated wavelet transform (QNDWT2D) to produce multiscale directional color-coupled coefficients, and extracting interpretable feature families (stain balance, wavelet energy, phase concentration, anisotropy, etc.) for four-class classification (normal, benign, in situ, invasive) on the BACH dataset via radial SVM with repeated nested cross-validation. It reports balanced recognition with errors mainly between adjacent classes and uses permutation importance to argue that directional, phase, anisotropy, scale, and amplitude groups reflect genuine quaternion multiscale geometry rather than global color.

Significance. If the quaternion coupling demonstrably improves over per-channel scalar wavelets while preserving interpretability and reproducibility without pretrained networks, the work supplies a useful baseline for computational pathology that ties features directly to histopathological properties such as nuclear density and glandular organization.

major comments (2)
  1. [Abstract / Permutation importance analysis] Abstract and results on permutation importance: the claim that the SVM 'draws on genuine quaternion and multiscale geometry rather than global color alone' is load-bearing for the central interpretation yet rests on an untested assumption; no ablation replaces QNDWT2D with three independent scalar NDWT2D runs on R/G/B (or decorrelated channels) while retaining identical feature families, SVM, and CV protocol. Equivalent performance from the scalar baseline would falsify the quaternion-specific contribution.
  2. [Evaluation section] Evaluation protocol: the abstract states repeated nested cross-validation but provides neither exact SVM radial-kernel hyperparameters (C, gamma), full per-fold confusion matrices, nor error bars on accuracy/F1; these omissions leave the 'balanced recognition' claim defensible but unverifiable at the level required for the reported class-wise error patterns.
minor comments (2)
  1. [Methods] Notation for quaternion encoding (pure quaternion field construction from RGB) and the precise definition of each feature family (e.g., 'phase concentration', 'color-axis geometry') should be given explicitly with equations in the methods section.
  2. [Figures/Tables] Figure captions and tables should include the number of repetitions in the nested CV and the precise train/test split sizes used on BACH.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the two major comments, which highlight opportunities to strengthen the evidence for the quaternion contribution and to improve reproducibility. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract / Permutation importance analysis] Abstract and results on permutation importance: the claim that the SVM 'draws on genuine quaternion and multiscale geometry rather than global color alone' is load-bearing for the central interpretation yet rests on an untested assumption; no ablation replaces QNDWT2D with three independent scalar NDWT2D runs on R/G/B (or decorrelated channels) while retaining identical feature families, SVM, and CV protocol. Equivalent performance from the scalar baseline would falsify the quaternion-specific contribution.

    Authors: We agree that the current permutation-importance analysis alone does not fully isolate the benefit of quaternion coupling from possible gains due to the multiscale directional features themselves. An explicit ablation that applies three independent scalar NDWT2D transforms to the R, G, and B channels (or to decorrelated channels) while keeping the identical feature families, radial SVM, and repeated nested CV protocol would directly test whether the quaternion formulation adds value. We will conduct this ablation on the BACH dataset and report the comparative accuracies, F1 scores, and confusion matrices in the revised manuscript. revision: yes

  2. Referee: [Evaluation section] Evaluation protocol: the abstract states repeated nested cross-validation but provides neither exact SVM radial-kernel hyperparameters (C, gamma), full per-fold confusion matrices, nor error bars on accuracy/F1; these omissions leave the 'balanced recognition' claim defensible but unverifiable at the level required for the reported class-wise error patterns.

    Authors: We acknowledge that the current manuscript does not report the exact radial-kernel hyperparameters selected by the inner loop of the nested CV, the full set of per-fold confusion matrices, or error bars on the performance metrics. In the revision we will add these details: the optimal (C, gamma) pairs, the complete per-fold confusion matrices (or their averages), and standard-error bars on accuracy and macro-F1 obtained across the outer CV repetitions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; feature construction and evaluation remain independent

full rationale

The paper encodes RGB images as quaternion fields, applies QNDWT2D to obtain coefficient fields, and constructs explicit feature families (stain balance, energy, phase concentration, anisotropy, etc.) directly from those coefficients. These descriptors are then fed to an SVM and assessed via repeated nested cross-validation on the external BACH dataset. No equation reduces a reported performance metric or importance ranking to a fitted parameter by construction, and no self-citation is shown to supply a uniqueness theorem or ansatz that the central claims rest upon. The derivation chain is therefore self-contained against the external benchmark.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard quaternion algebra and wavelet theory plus the domain assumption that the chosen features map meaningfully to tissue properties; no new entities are postulated and the only free parameters are those implicit in SVM hyperparameter tuning.

free parameters (1)
  • SVM radial kernel hyperparameters
    The radial kernel SVM requires tuning of C and gamma during nested cross-validation; these are fitted to the BACH data.
axioms (2)
  • standard math Quaternion multiplication and conjugation rules support a well-defined nondecimated wavelet transform in 2D
    Invoked when defining QNDWT2D on pure quaternion image fields.
  • domain assumption The listed feature families (energy, phase concentration, anisotropy, etc.) correspond to histopathological properties
    Stated when connecting descriptors to nuclear density and glandular organization.

pith-pipeline@v0.9.1-grok · 5830 in / 1508 out tokens · 58188 ms · 2026-07-03T03:02:16.649034+00:00 · methodology

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

47 extracted references

  1. [1]

    Proceedings of the Royal Irish Academy , year =

    Hamilton, William Rowan , title =. Proceedings of the Royal Irish Academy , year =

  2. [2]

    Journal of Mathematical Imaging and Vision , volume =

    Bayro-Corrochano, Eduardo , title =. Journal of Mathematical Imaging and Vision , volume =

  3. [3]

    , title =

    Chan, Wei Lu and Choi, Hyeokho and Baraniuk, Richard G. , title =. IEEE Transactions on Image Processing , volume =. 2008 , doi =

    2008

  4. [4]

    Hypercomplex Spectral Signal Representations for the Processing and Analysis of Images , school =

    B. Hypercomplex Spectral Signal Representations for the Processing and Analysis of Images , school =

  5. [5]

    Fletcher, Peter , title =

  6. [6]

    , title =

    Fletcher, Peter and Sangwine, Stephen J. , title =. Signal Processing , volume =. 2017 , doi =

    2017

  7. [7]

    Quaternionic Wavelets for Texture Classification , journal =

    Soulard, Rapha. Quaternionic Wavelets for Texture Classification , journal =. 2011 , doi =

    2011

  8. [8]

    International Journal of Electronics and Communications (AE

    Gai, Shan and Yang, Guowei and Zhang, Sheng , title =. International Journal of Electronics and Communications (AE. 2013 , doi =

    2013

  9. [9]

    Mathematical Problems in Engineering , volume =

    Yin, Ming and Liu, Wei and Shui, Jun and Wu, Jiangmin , title =. Mathematical Problems in Engineering , volume =. 2012 , doi =

    2012

  10. [10]

    Neurocomputing , volume =

    Sigillo, Luigi and Grassucci, Eleonora and Uncini, Aurelio and Comminiello, Danilo , title =. Neurocomputing , volume =. 2025 , doi =

    2025

  11. [11]

    and Boucheron, Laura E

    Gurcan, Metin N. and Boucheron, Laura E. and Can, Ahmet and Madabhushi, Anant and Rajpoot, Nasir M. and Yener, B. Histopathological Image Analysis: A Review , journal =. 2009 , doi =

    2009

  12. [12]

    Veta, Mitko and Pluim, Josien P. W. and van Diest, Paul J. and Viergever, Max A. , title =. IEEE Transactions on Biomedical Engineering , volume =. 2014 , doi =

    2014

  13. [13]

    and Oliveira, Luiz S

    Spanhol, Fabio A. and Oliveira, Luiz S. and Petitjean, Cl. A Dataset for Breast Cancer Histopathological Image Classification , journal =. 2016 , doi =

    2016

  14. [14]

    BACH: Grand Challenge on Breast Cancer Histology Images , booktitle =

    Aresta, Guilherme and Ara. BACH: Grand Challenge on Breast Cancer Histology Images , booktitle =. 2019 , doi =

    2019

  15. [15]

    Litjens, Geert and Kooi, Thijs and Bejnordi, Babak Ehteshami and Setio, Arnaud Arindra Adiyoso and Ciompi, Francesco and Ghafoorian, Mohsen and van der Laak, Jeroen A. W. M. and van Ginneken, Bram and S. A Survey on Deep Learning in Medical Image Analysis , journal =. 2017 , doi =

    2017

  16. [16]

    Computational and Structural Biotechnology Journal , volume =

    Komura, Daisuke and Ishikawa, Shumpei , title =. Computational and Structural Biotechnology Journal , volume =. 2018 , doi =

    2018

  17. [17]

    , title =

    Donoho, David L. , title =. IEEE Transactions on Information Theory , volume =

  18. [18]

    Medical Image Analysis , volume =

    Aresta, Guilherme and Ara. Medical Image Analysis , volume =. 2019 , doi =

    2019

  19. [19]

    Dataset: Grand Challenge on Breast Cancer Histology Images , year =

  20. [20]

    2019 , publisher =

    Pol. 2019 , publisher =

    2019

  21. [21]

    2018 , howpublished =

    Vesal, Sulaiman and Ravikumar, Nishant and Davari, AmirAbbas and Ellmann, Stephan and Maier, Andreas , title =. 2018 , howpublished =

    2018

  22. [22]

    International Conference on Image Analysis and Recognition , pages=

    Classification of Breast Cancer Histology Images Using Transfer Learning , author=. International Conference on Image Analysis and Recognition , pages=. 2018 , organization=

    2018

  23. [23]

    2018 , howpublished =

    Nazeri, Kamyar and Aminpour, Azad and Ebrahimi, Mehran , title =. 2018 , howpublished =

    2018

  24. [24]

    International Conference on Image Analysis and Recognition , pages=

    Two-Stage Convolutional Neural Network for Breast Cancer Histology Image Classification , author=. International Conference on Image Analysis and Recognition , pages=. 2018 , organization=

    2018

  25. [25]

    Medical Image Analysis , volume=

    BACH: Grand challenge on breast cancer histology images , author=. Medical Image Analysis , volume=. 2019 , publisher=

    2019

  26. [26]

    PLOS ONE , volume=

    Classification of breast cancer histology images using convolutional neural networks , author=. PLOS ONE , volume=. 2017 , publisher=

    2017

  27. [27]

    Computers in Biology and Medicine , volume=

    Automatic classification of tissue malignancy for breast carcinoma diagnosis , author=. Computers in Biology and Medicine , volume=. 2018 , publisher=

    2018

  28. [28]

    International Conference on Image Analysis and Recognition , pages=

    Ensemble Network for Region Identification in Breast Histopathology Slides , author=. International Conference on Image Analysis and Recognition , pages=. 2018 , organization=

    2018

  29. [29]

    1992 , doi =

    Daubechies, Ingrid , title =. 1992 , doi =

    1992

  30. [30]

    A Theory for Multiresolution Signal Decomposition: The Wavelet Representation , journal =

    Mallat, St. A Theory for Multiresolution Signal Decomposition: The Wavelet Representation , journal =. 1989 , doi =

    1989

  31. [31]

    and Silverman, Bernard W

    Nason, Guy P. and Silverman, Bernard W. , title =. Wavelets and Statistics , editor =. 1995 , doi =

    1995

  32. [32]

    , title =

    Starck, Jean-Luc and Murtagh, Fionn and Fadili, Jalal M. , title =

  33. [33]

    and Shanmugam, K

    Haralick, Robert M. and Shanmugam, K. and Dinstein, Its'hak , title =. IEEE Transactions on Systems, Man, and Cybernetics , volume =. 1973 , doi =

    1973

  34. [34]

    IEEE Transactions on Image Processing , volume =

    Unser, Michael , title =. IEEE Transactions on Image Processing , volume =. 1995 , doi =

    1995

  35. [35]

    , title =

    Shannon, Claude E. , title =. The Bell System Technical Journal , volume =. 1948 , doi =

    1948

  36. [36]

    , title =

    Fisher, Nicholas I. , title =. 1993 , doi =

    1993

  37. [37]

    and Jupp, Peter E

    Mardia, Kanti V. and Jupp, Peter E. , title =. 2000 , doi =

    2000

  38. [38]

    and Johnston, Dennis A

    Ruifrok, Arnout C. and Johnston, Dennis A. , title =. Analytical and Quantitative Cytology and Histology , volume =. 2001 , pmid =

    2001

  39. [39]

    Machine Learning , volume =

    Cortes, Corinna and Vapnik, Vladimir , title =. Machine Learning , volume =. 1995 , doi =

    1995

  40. [40]

    Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond , publisher =

    Sch. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond , publisher =. 2002 , isbn =

    2002

  41. [41]

    IEEE Transactions on Neural Networks , volume =

    Hsu, Chih-Wei and Lin, Chih-Jen , title =. IEEE Transactions on Neural Networks , volume =. 2002 , doi =

    2002

  42. [42]

    and Talbot, Nicola L

    Cawley, Gavin C. and Talbot, Nicola L. C. , title =. Journal of Machine Learning Research , volume =

  43. [43]

    Machine Learning , volume =

    Breiman, Leo , title =. Machine Learning , volume =. 2001 , doi =

    2001

  44. [44]

    Permutation Importance: A Corrected Feature Importance Measure , journal =

    Altmann, Andr. Permutation Importance: A Corrected Feature Importance Measure , journal =. 2010 , doi =

    2010

  45. [45]

    Analytical Cellular Pathology , volume =

    Hwang, Hae-Gil and Choi, Hyun-Ju and Lee, Byeong-Il and Yoon, Hye-Kyoung and Nam, Sang-Hee and Choi, Heung-Kook , title =. Analytical Cellular Pathology , volume =. 2005 , doi =

    2005

  46. [46]

    Frontiers in Oncology , volume =

    Chaddad, Ahmad and Daniel, Paul and Niazi, Tamim , title =. Frontiers in Oncology , volume =. 2018 , doi =

    2018

  47. [47]

    Laboratory Investigation , volume =

    Lu, Cheng and Romo-Bucheli, David and Wang, Xiangxue and Janowczyk, Andrew and Ganesan, Shridar and Gilmore, Hannah and Rimm, David and Madabhushi, Anant , title =. Laboratory Investigation , volume =. 2018 , doi =

    2018