pith. sign in

arxiv: 2604.19823 · v1 · submitted 2026-04-20 · 💻 cs.CV · cs.AI· cs.LG

Rabies diagnosis in low-data settings: A comparative study on the impact of data augmentation and transfer learning

Khalil Akremi , Mariem Handous , Zied Bouslama , Farah Bassalah , Maryem Jebali , Mariem Hanachi , Ines Abdeljaoued-Tej This is my paper

Pith reviewed 2026-05-10 05:35 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords rabies diagnosistransfer learningdata augmentationfluorescent imagesdeep learningmedical imaginglow-data settingscomputer vision
0
0 comments X

The pith

EfficientNetB0 with geometric and color augmentations achieves optimal performance classifying rabies fluorescent images.

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

The paper investigates whether transfer learning and data augmentation can automate interpretation of fluorescent microscope images for rabies diagnosis when only a small number of examples exist. This would matter in regions where rabies surveillance depends on accurate results but skilled microscopists are scarce and annual sample volumes are low. Four architectures were compared on 155 images using three augmentation strategies, with models selected through stratified cross-validation. TrivialAugmentWide preserved the critical glowing patterns needed for positive cases better than alternatives, leading to the best results from EfficientNetB0 on cropped images. The authors also released an online tool to make the system immediately usable in practice.

Core claim

The EfficientNetB0 model utilizing Geometric and Color augmentation and selected through stratified 3-fold cross-validation achieved optimal classification performance on cropped images, while TrivialAugmentWide was the most effective augmentation technique as it preserved critical fluorescent patterns while improving model robustness on the 155-image dataset.

What carries the argument

Transfer learning applied to pre-trained models on cropped fluorescent images, with data augmentation strategies compared to enhance generalization despite limited samples.

If this is right

  • Fast and reliable rabies detection becomes possible without requiring constant expert interpretation on site.
  • Deep learning proves viable for automating this diagnostic task even under constraints of small and imbalanced data.
  • The deployed online tool provides immediate practical access for users in affected areas.
  • The pipeline creates a reusable framework for similar medical imaging applications with limited training data.

Where Pith is reading between the lines

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

  • The same combination of transfer learning and carefully chosen augmentations may support automated diagnosis for other fluorescence-based tests in comparable low-data environments.
  • Augmentation methods that avoid distorting key visual markers such as fluorescent signals appear especially important when sample sizes are small.
  • Testing the system on images from a wider range of laboratories would provide a direct check on how well the reported performance holds outside the original collection.

Load-bearing premise

A dataset of 155 images with noted class imbalance is representative enough for the trained models to generalize reliably to varied real-world conditions in low-resource laboratories.

What would settle it

If accuracy falls substantially when the best model is evaluated on a new collection of fluorescent images gathered from multiple independent low-resource labs in different regions, the generalization claim would not hold.

Figures

Figures reproduced from arXiv: 2604.19823 by Farah Bassalah, Ines Abdeljaoued-Tej, Khalil Akremi, Mariem Hanachi, Mariem Handous, Maryem Jebali, Zied Bouslama.

Figure 1
Figure 1. Figure 1: Comparison of original (left) and YOLO-annotated (right) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Impact of image preprocessing on model performance [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Grad-CAM Model Interpretability: Visualization of learned features highlighting rabies-infected regions [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Model performance across all aonfigurations: Accuracy evolution during training [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training and validation Loss curves: Overfitting detection across model architectures and augmentation strategies [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
read the original abstract

Rabies remains a major public health concern across many African and Asian countries, where accurate diagnosis is critical for effective epidemiological surveillance. The gold standard diagnostic methods rely heavily on fluorescence microscopy, necessitating skilled laboratory personnel for the accurate interpretation of results. Such expertise is often scarce, particularly in regions with low annual sample volumes. This paper presents an automated, AI-driven diagnostic system designed to address these challenges. We developed a robust pipeline utilizing fluorescent image analysis through transfer learning with four deep learning architectures: EfficientNetB0, EfficientNetB2, VGG16, and Vision Transformer (ViTB16). Three distinct data augmentation strategies were evaluated to enhance model generalization on a dataset of 155 microscopic images (123 positive and 32 negative). Our results demonstrate that TrivialAugmentWide was the most effective augmentation technique, as it preserved critical fluorescent patterns while improving model robustness. The EfficientNetB0 model, utilizing Geometric & Color augmentation and selected through stratified 3fold cross-validation, achieved optimal classification performance on cropped images. Despite constraints posed by class imbalance and a limited dataset size, this work confirms the viability of deep learning for automating rabies diagnosis. The proposed method enables fast and reliable detection with significant potential for further optimization. An online tool was deployed to facilitate practical access, establishing a framework for future medical imaging applications. This research underscores the potential of optimized deep learning models to transform rabies diagnostics and improve public health outcomes.

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

3 major / 2 minor

Summary. The manuscript presents an empirical comparison of four transfer learning architectures (EfficientNetB0, EfficientNetB2, VGG16, ViT-B16) paired with three data augmentation strategies for binary classification of rabies in a dataset of 155 fluorescent microscopy images (123 positive, 32 negative). Using stratified 3-fold cross-validation, it identifies EfficientNetB0 with Geometric & Color augmentation as optimal on cropped images and TrivialAugmentWide as the most effective augmentation technique overall. The work concludes that the approach demonstrates the viability of deep learning for automated rabies diagnosis in low-data settings and deploys an online tool for practical use.

Significance. If the internal performance estimates prove representative, the study could support AI tools for rabies diagnosis in resource-limited laboratories where expert microscopists are scarce. The comparative evaluation across architectures and augmentations, together with the deployed online tool, provides a concrete starting point for further medical imaging applications in low-resource contexts. The small dataset size and evaluation design, however, limit the strength of claims about real-world reliability and generalization.

major comments (3)
  1. [Methods] Methods section: The stratified 3-fold cross-validation is described without details on class-imbalance mitigation (e.g., loss weighting, oversampling, or threshold adjustment) or per-fold sensitivity/specificity values with confidence intervals. With only ~10–11 negative samples per fold, this omission makes it impossible to evaluate whether the reported optimality of EfficientNetB0 is robust or an artifact of the particular split.
  2. [Results] Results section: The central claim that EfficientNetB0 with Geometric & Color augmentation yields 'optimal classification performance' and that TrivialAugmentWide is 'most effective' rests entirely on internal 3-fold CV. No external held-out test set from different microscopes, laboratories, or geographies is reported, so the generalization argument for diverse low-resource settings lacks direct empirical support.
  3. [Discussion] Discussion/Conclusion: The statements that the method enables 'fast and reliable detection' and confirms 'viability' for real-world rabies diagnosis are not proportionate to the evidence; the 155-image dataset with acknowledged imbalance and no multi-center validation leaves open the possibility that performance reflects dataset-specific fluorescent artifacts rather than transferable pattern detection.
minor comments (2)
  1. [Abstract] Abstract: The claim of 'optimal classification performance' is stated without any numerical results (accuracy, sensitivity, specificity, or AUC), which weakens the abstract's utility as a standalone summary.
  2. [Methods] The manuscript would benefit from explicit reporting of the exact cropping coordinates or procedure and from qualitative examples showing how cropping affects fluorescent inclusion bodies.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive feedback on our manuscript. We have addressed each of the major comments in detail below and plan to revise the paper accordingly to improve clarity and temper our claims where appropriate.

read point-by-point responses

  1. Referee: [Methods] Methods section: The stratified 3-fold cross-validation is described without details on class-imbalance mitigation (e.g., loss weighting, oversampling, or threshold adjustment) or per-fold sensitivity/specificity values with confidence intervals. With only ~10–11 negative samples per fold, this omission makes it impossible to evaluate whether the reported optimality of EfficientNetB0 is robust or an artifact of the particular split.

    Authors: We acknowledge the need for greater transparency in our evaluation protocol. Our stratified 3-fold cross-validation ensured that each fold maintained the overall class distribution, but we did not implement additional imbalance mitigation strategies such as weighted loss or oversampling. We will revise the Methods section to document this explicitly. Additionally, we will report the per-fold performance metrics, including sensitivity and specificity, along with 95% confidence intervals estimated using bootstrap methods. This will enable assessment of the robustness of the EfficientNetB0 results across different splits, despite the small number of negative samples per fold. revision: yes

  2. Referee: [Results] Results section: The central claim that EfficientNetB0 with Geometric & Color augmentation yields 'optimal classification performance' and that TrivialAugmentWide is 'most effective' rests entirely on internal 3-fold CV. No external held-out test set from different microscopes, laboratories, or geographies is reported, so the generalization argument for diverse low-resource settings lacks direct empirical support.

    Authors: We agree that the absence of an external test set limits the strength of generalization claims. The study focuses on a comparative analysis within a constrained low-data regime using internal validation. We will modify the Results section to present the findings as internal performance estimates and will add explicit caveats regarding generalization to new settings. While we cannot add an external test set without new data, the online tool we deployed allows for prospective testing on external images, which we will highlight as a means for future validation. revision: partial

  3. Referee: [Discussion] Discussion/Conclusion: The statements that the method enables 'fast and reliable detection' and confirms 'viability' for real-world rabies diagnosis are not proportionate to the evidence; the 155-image dataset with acknowledged imbalance and no multi-center validation leaves open the possibility that performance reflects dataset-specific fluorescent artifacts rather than transferable pattern detection.

    Authors: We accept that some language in the Discussion and Conclusion may overstate the current evidence. We will revise these sections to use more measured terms, such as 'suggests the potential for' and 'provides initial support for the viability', and will include a new paragraph on limitations. This will address the small sample size, class imbalance, single-source data, and the possibility of capturing dataset-specific artifacts, while outlining plans for multi-center validation in future work. revision: yes

standing simulated objections not resolved
  • Lack of external validation data from diverse sources, which cannot be provided without acquiring additional images from different laboratories or geographies.

Circularity Check

0 steps flagged

No circularity: purely empirical ML comparison with direct evaluation on fixed dataset

full rationale

The paper reports results from training and evaluating four transfer learning architectures (EfficientNetB0, EfficientNetB2, VGG16, ViTB16) plus three augmentation strategies on a fixed set of 155 images using stratified 3-fold cross-validation. No mathematical derivations, first-principles predictions, fitted parameters renamed as outputs, or self-citations that bear the central claim are present. Performance numbers and model optimality statements arise directly from standard training/evaluation loops on the provided data rather than any reduction by construction. The study is self-contained as an experimental benchmark.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the representativeness of the 155-image dataset and the assumption that ImageNet-pretrained models transfer effectively to fluorescent microscopy without domain-specific adaptation beyond standard fine-tuning.

free parameters (2)
  • Augmentation strategy selection
    Choice among three augmentation methods and their parameters was determined by performance on the small dataset.
  • Model architecture and cropping decisions
    Selection of EfficientNetB0 and image cropping was tuned via cross-validation on the given data.
axioms (1)
  • domain assumption Transfer learning from natural-image pretraining generalizes to fluorescent rabies microscopy images
    Invoked by applying EfficientNet, VGG, and ViT models without additional domain adaptation justification.

pith-pipeline@v0.9.0 · 5592 in / 1517 out tokens · 46936 ms · 2026-05-10T05:35:27.743281+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

28 extracted references · 1 canonical work pages

  1. [1]

    Magalhães, and Sunanda J

    Abràmoff, Michael D., Paulo J. Magalhães, and Sunanda J. Ram. ”Image processing with ImageJ. ” Biophotonics interna- tional 11.7 (2004): 36-42

    2004

  2. [2]

    ”Image classification using deep learn- ing: A comparative study of vgg-16, inceptionv3 and efficient- net b7 models

    Aggarwal, Shivam, et al. ”Image classification using deep learn- ing: A comparative study of vgg-16, inceptionv3 and efficient- net b7 models. ” 2023 3rd International Conference on Ad- vance Computing and Innovative Technologies in Engineering (ICACITE). IEEE, 2023

    2023

  3. [3]

    ”RoboFlow: A flow-based visual programming language for mobile manipulation tasks

    Alexandrova, Sonya, Zachary Tatlock, and Maya Cakmak. ”RoboFlow: A flow-based visual programming language for mobile manipulation tasks. ” 2015 IEEE international conference on robotics and automation (ICRA). IEEE, 2015

    2015

  4. [4]

    ”A Child Surviving Rabies in Tunisia: A Case Report

    Asma, Ben Halima, et al. ”A Child Surviving Rabies in Tunisia: A Case Report. ” Indian Journal of Pediatrics 91.3 (2024): 308- 308

    2024

  5. [5]

    ”QuPath: Open source software for digital pathology image analysis

    Bankhead, Peter, et al. ”QuPath: Open source software for digital pathology image analysis. ” Scientific reports 7.1 (2017): 1-7

    2017

  6. [6]

    ”Molecular epidemiology of rabies in wild canidae in Tunisia

    Bouslama, Zied, et al. ”Molecular epidemiology of rabies in wild canidae in Tunisia. ” Viruses 13.12 (2021): 2473

    2021

  7. [7]

    Development of a fuores- cent antibody virus neutralisation test (F A VN test) for the quantitation of rabies-neutralising antibody

    Cliquet F, Aubert M, Sagné L. Development of a fuores- cent antibody virus neutralisation test (F A VN test) for the quantitation of rabies-neutralising antibody. J Immunol Meth- ods. 1998;212:79–87. https://doi.org/10.1016/S0022-1759(97) 00212-3

    work page doi:10.1016/s0022-1759(97 1998

  8. [8]

    ”Medical image data augmentation: techniques, comparisons and interpretations

    Goceri, Evgin. ”Medical image data augmentation: techniques, comparisons and interpretations. ” Artificial intelligence review 56.11 (2023): 12561-12605

    2023

  9. [9]

    ”Alpha-divergence for classification, indexing and retrieval

    Hero, Alfred O., et al. ”Alpha-divergence for classification, indexing and retrieval. ” Communication and Signal Processing Laboratory, Technical Report CSPL-328, U. Mich (2001)

    2001

  10. [10]

    Huang, Liang-Kai, and Mao-Jiun J. Wang. ”Image thresholding by minimizing the measures of fuzziness. ” Pattern recognition 28.1 (1995): 41-51

    1995

  11. [11]

    Sahoo, and Andrew KC Wong

    Kapur, Jagat Narain, Prasanna K. Sahoo, and Andrew KC Wong. ”A new method for gray-level picture thresholding using the entropy of the histogram. ” Computer vision, graphics, and image processing 29.3 (1985): 273-285

    1985

  12. [12]

    ”Distinct Views Improve Generalization and Robustness: Combinations of Augmentations With Different Features

    Kim, Keon, Hyun woo Kim, and Yong Suk Choi. ”Distinct Views Improve Generalization and Robustness: Combinations of Augmentations With Different Features. ” IEEE Access (2025)

    2025

  13. [13]

    ”EfficientNet

    Koonce, Brett. ”EfficientNet. ” Convolutional neural networks with swift for Tensorflow: image recognition and dataset cate- gorization. Berkeley, CA: Apress, 2021. 109-123

    2021

  14. [14]

    Li, Chun Hung, and C. K. Lee. ”Minimum cross entropy thresholding. ” Pattern recognition 26.4 (1993): 617-625

    1993

  15. [15]

    ”CellProfiler 3.0: Next-generation image processing for biology

    McQuin, Claire, et al. ”CellProfiler 3.0: Next-generation image processing for biology. ” PLoS biology 16.7 (2018): e2005970

    2018

  16. [16]

    ”Trivialaugment: Tuning- free yet state-of-the-art data augmentation

    Müller, Samuel G., and Frank Hutter. ”Trivialaugment: Tuning- free yet state-of-the-art data augmentation. ” Proceedings of the IEEE/CVF international conference on computer vision. 2021

    2021

  17. [17]

    Otsu, N. (1975). A threshold selection method from gray-level histograms. Automatica, 11(285-296), 23-27

    1975

  18. [18]

    Otsu, N. (1979). ”A Threshold Selection Method from Gray- Level Histograms. ” IEEE Transactions on Systems, Man, and Cybernetics 9(1): 62–66

    1979

  19. [19]

    Ridler, Thomas Wilhelm, and S. Calvard. ”Picture thresholding using an iterative selection method. ” IEEE Trans. Syst. Man Cybern 8.8 (1978): 630-632

    1978

  20. [20]

    Laboratory tech- niques in rabies

    Rupprecht CE, Fooks AR, Abela-Ridder B. Laboratory tech- niques in rabies. 5th ed. 2018. p. 302

    2018

  21. [21]

    Khoshgoftaar

    Shorten, Connor, and Taghi M. Khoshgoftaar. ”A survey on image data augmentation for deep learning. ” Journal of big data 6.1 (2019): 1-48

    2019

  22. [22]

    ”A review on yolov8 and its advancements

    Sohan, Mupparaju, Thotakura Sai Ram, and Ch Venkata Rami Reddy. ”A review on yolov8 and its advancements. ” International Conference on Data Intelligence and Cognitive Informatics. Springer, Singapore, 2024

    2024

  23. [23]

    ”Transfer learning using vgg-16 with deep convolutional neural network for classifying images

    Tammina, Srikanth. ”Transfer learning using vgg-16 with deep convolutional neural network for classifying images. ” Interna- tional Journal of Scientific and Research Publications (IJSRP) 9.10 (2019): 143-150

    2019

  24. [24]

    Laboratory techniques in rabies, volume 1, 5th ed

    World Health Organization, Rupprecht, Charles E, Fooks, Anthony R & Abela-Ridder, Bernadette (2018). Laboratory techniques in rabies, volume 1, 5th ed. World Health Organi- zation. https://iris.who.int/handle/10665/310836. Licence: CC BY-NC-SA 3.0 IGO

    2018

  25. [25]

    World Organisation for Animal Health

    WOAH, Terrestrial Manual: Chapter 3.1.18 — Rabies (In- fection with rabies virus and other lyssaviruses), version adopted May 2023. World Organisation for Animal Health. [On- line]. A vailable: https://www.woah.org/fileadmin/Home/eng/ Health_standards/tahm/3.01.18_RABIES.pdf

    2023

  26. [26]

    Rabies (infection with Rabies virus and other lyssaviruses)

    World Organization of Animal Health. Rabies (infection with Rabies virus and other lyssaviruses). OIE Terrestrial Man- ual; 2018. p. 34. https://www.woah.org/fleadmin/Home/eng/ Health_standards/tahm/3.01.18_RABIES.pdf

    2018

  27. [27]

    25, 2025

    World Health Organization, ”Rabies — Epidemiology and burden of disease,” Control of Neglected Tropical Diseases, accessed: Sep. 25, 2025. [Online]. A vailable: https: //www.who.int/teams/control-of-neglected-tropical-diseases/ rabies/epidemiology-and-burden

    2025

  28. [28]

    Rogers, and Samuel A

    Zack, Gregory W., William E. Rogers, and Samuel A. Latt. ”Au- tomatic measurement of sister chromatid exchange frequency. ” Journal of Histochemistry & Cytochemistry 25.7 (1977): 741- 753

    1977