Beyond Nearest Neighbor Interpolation in Data Augmentation

Olivier Rukundo

arxiv: 2504.01527 · v3 · submitted 2025-04-02 · 💻 cs.CV · eess.IV

Beyond Nearest Neighbor Interpolation in Data Augmentation

Olivier Rukundo This is my paper

Pith reviewed 2026-05-22 22:08 UTC · model grok-4.3

classification 💻 cs.CV eess.IV

keywords data augmentationnearest neighbor interpolationmedical image segmentationconvolutional neural networksgeometric transformationclass filteringlow-pass filteringaugmented training data

0 comments

The pith

Replacing nearest neighbor interpolation with alternative methods and a mean-based class filter improves performance in medical image segmentation.

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

The paper aims to establish that nearest neighbor interpolation in data augmentation for labeled images risks worsening pixel-level annotation errors and degrading high-frequency details through low-pass filtering. The author addresses this by modifying convolutional neural network data transformation functions to incorporate a modified geometric transformation, drop reliance on nearest neighbor interpolation, and add a mean-based class filtering mechanism for handling undefined labels with other interpolators. An offline data augmentation pipeline generates the specific augmented sets needed to measure these effects. Experiments on three medical image segmentation datasets plus the XBAT+ datasets show gains across quantitative metrics.

Core claim

The central claim is that modifying the data transformation functions by incorporating a modified geometric transformation function, removing reliance on nearest neighbor interpolation, and integrating a mean-based class filtering mechanism enables the use of alternative interpolation algorithms. This avoids the risk of undefined categorical labels and the exacerbation of pixel level annotation errors while reducing degradation of high-frequency structural details within annotated regions of interest.

What carries the argument

Modified geometric transformation function combined with mean-based class filtering mechanism to support alternative interpolations without nearest neighbor reliance.

If this is right

Performance gains across multiple quantitative metrics on medical image segmentation datasets and XBAT+.
Reduced risk of exacerbating pixel level annotation errors in the augmented training data.
Reduced degradation of high-frequency structural details within annotated regions of interest.
Quantitative assessment of interpolation-specific low-pass filtering effects on augmented training data.

Where Pith is reading between the lines

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

The offline pipeline approach could be adapted to compare interpolation effects in non-medical computer vision tasks.
The filtering step might interact with other augmentation strategies such as random cropping or color jitter.
Similar modifications could be tested on regression tasks where label continuity matters more than categorical labels.

Load-bearing premise

The mean-based class filtering mechanism handles undefined categorical labels without introducing new biases or errors while preserving high-frequency details.

What would settle it

If the proposed modifications produce no gains or lower segmentation accuracy than nearest neighbor interpolation on the same medical datasets, the performance improvement claim would not hold.

Figures

Figures reproduced from arXiv: 2504.01527 by Olivier Rukundo.

read the original abstract

Avoiding the risk of undefined categorical labels using nearest neighbor interpolation overlooks the risk of exacerbating pixel level annotation errors in augmented training data. Additionally, the inherent low pass filtering effects of interpolation algorithms exacerbate the risk of degrading high frequency structural details within annotated regions of interest. To avoid these risks, the author modified convolutional neural networks data transformation functions by incorporating a modified geometric transformation function, removing reliance on nearest neighbor interpolation, and integrating a mean-based class filtering mechanism to handle undefined categorical labels with alternative interpolation algorithms. The author also implemented an offline data augmentation pipeline to generate interpolation specific augmented training data, enabling quantitative assessment of interpolation specific low pass filtering effects on augmented training data. Experimental evaluation on three medical image segmentation datasets and the XBAT+ datasets demonstrated performance gains across multiple quantitative metrics.

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.

Desk Editor's Note private letter to a colleague

The paper flags a real issue with nearest-neighbor label interpolation in augmentation but the mean-based fix risks reintroducing the exact smoothing it aims to avoid.

read the letter

The core observation here is practical: nearest-neighbor interpolation for categorical labels in medical segmentation augmentation can worsen existing annotation errors, while other interpolation methods introduce low-pass filtering that blurs high-frequency details. The author responds by altering the geometric transformation functions, dropping nearest-neighbor reliance, and adding a mean-based class filter for undefined labels, plus an offline pipeline that generates and quantifies interpolation-specific augmented data. That pipeline is a straightforward way to measure the low-pass effects on training sets, which is useful for anyone running controlled augmentation experiments. The work stays within established computer vision practices rather than introducing a new framework. The experimental claim is that the changes produced gains across quantitative metrics on three medical segmentation datasets plus XBAT+, but the abstract supplies no numbers, baselines, variance, or statistical tests, so the strength of that support cannot be judged from the summary alone. The bigger concern is the mean-based filter itself. If it computes per-pixel or per-channel averages to resolve undefined labels, it functions as a low-pass operation precisely at object boundaries, which directly conflicts with the stated goal of preserving structural detail. The abstract gives no derivation, masking rule, or threshold that would prevent this, so the central mechanism needs explicit verification in the full methods. This paper is aimed at researchers already working on data augmentation for pixel-level medical tasks who want incremental pipeline tweaks. It is coherent on its own terms and engages the literature on interpolation artifacts, so it clears the bar for a serious referee even if the results turn out modest. I would send it to review to check the filter implementation and the actual metrics rather than desk-reject it.

Referee Report

1 major / 1 minor

Summary. The paper claims that nearest-neighbor interpolation for categorical labels in data augmentation risks exacerbating pixel-level annotation errors and that interpolation algorithms introduce low-pass filtering that degrades high-frequency details in regions of interest. It proposes modifying CNN geometric transformation functions to remove reliance on nearest-neighbor interpolation, integrating a mean-based class filtering mechanism to handle undefined labels with alternative interpolators, and implementing an offline augmentation pipeline to generate interpolation-specific augmented data for quantitative assessment. Experiments on three medical image segmentation datasets and the XBAT+ datasets are reported to demonstrate performance gains across multiple quantitative metrics.

Significance. If the mean-based filtering mechanism preserves high-frequency structural details without introducing boundary biases and the reported gains are robust, the work could offer a practical improvement to data augmentation pipelines in medical segmentation by enabling higher-quality interpolators. The offline pipeline for isolating interpolation effects is a methodological strength that supports reproducibility and targeted analysis.

major comments (1)

[Abstract] Abstract (and methods description): the mean-based class filtering mechanism is presented only at high level. If it computes per-pixel means across transformed class channels to resolve undefined labels, it applies a low-pass operation at object boundaries; this directly risks counteracting the stated goal of avoiding low-pass effects. No derivation, pseudocode, thresholding rule, or masking procedure is supplied to demonstrate how boundary smoothing is prevented while still handling undefined labels.

minor comments (1)

The abstract states that experiments demonstrated performance gains across multiple quantitative metrics but supplies no numerical values, baseline comparisons, dataset sizes, or statistical tests, which limits immediate assessment of the strength of the empirical support.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and detailed comments on the manuscript. We address the major comment below and will revise the paper accordingly.

read point-by-point responses

Referee: [Abstract] Abstract (and methods description): the mean-based class filtering mechanism is presented only at high level. If it computes per-pixel means across transformed class channels to resolve undefined labels, it applies a low-pass operation at object boundaries; this directly risks counteracting the stated goal of avoiding low-pass effects. No derivation, pseudocode, thresholding rule, or masking procedure is supplied to demonstrate how boundary smoothing is prevented while still handling undefined labels.

Authors: We agree that the mean-based class filtering mechanism is described only at a high level and that the current manuscript provides no derivation, pseudocode, thresholding rule, or masking procedure. This omission leaves open the possibility that the approach could introduce low-pass effects at boundaries, which would undermine the paper's central claim. In the revised version we will expand the methods section with a full mathematical derivation of the filtering step, pseudocode, explicit thresholding and masking rules, and an analysis showing how boundary smoothing is avoided while still resolving undefined labels after non-nearest-neighbor interpolation. revision: yes

Circularity Check

0 steps flagged

No circularity; experimental claims rest on external datasets and independent evaluation

full rationale

The paper describes a methodological change to data augmentation (modified geometric transforms plus mean-based class filtering to avoid nearest-neighbor interpolation) and reports performance gains on three medical segmentation datasets plus XBAT+. No equations, fitted parameters renamed as predictions, self-citations used as load-bearing uniqueness theorems, or ansatzes smuggled via prior work appear in the abstract or described pipeline. The evaluation is against external benchmarks, satisfying the self-contained criterion. No load-bearing step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient detail to identify specific free parameters, axioms, or invented entities; review limited to abstract only.

pith-pipeline@v0.9.0 · 5647 in / 1211 out tokens · 56593 ms · 2026-05-22T22:08:40.969900+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

INTRODUCTION Data augmentation operation involve s many techniques that operate at the data level rather than at the convolutional neural network (CNN) architectural level [1]. In the context of semantic segmentation with deep learning , this operation aims at increasing both training sets images and corresponding masks to help the CNN architecture of int...

work page
[2]

Datasets Three datasets were downloaded from Kaggle, a large platform hosting community -shared models, datasets, and codes [9]

MATERIALS AND METHODS A. Datasets Three datasets were downloaded from Kaggle, a large platform hosting community -shared models, datasets, and codes [9]. While selecting these datasets, the author focused on the diversity in digital imaging modalities and task of interest . The first dataset consists of fifty-five histology breast light- microscopy images...

work page
[3]

Selected Evaluation Metrics To evaluate the quality of predictions made by CNN architectures under various interpolation settings , Accuracy (Eq

EXPERIMENTAL RESULTS A. Selected Evaluation Metrics To evaluate the quality of predictions made by CNN architectures under various interpolation settings , Accuracy (Eq. 4), IoU (Eq. 5), and meanBFScore (Eq. 6 and Eq. 7), as well as the Dice score (Eq. 8), were used. These metrics were chosen due to their well -established suitability for assessing segmen...

work page
[4]

In Figure 5, in terms of Accuracy, Dice score, and IoU, the SEGNET's BIC_BIC were not ranked the first, at the exception of meanBFScore

DISCUSSIONS As can be seen in Figure 4, in terms of Accuracy, Dice score, IoU and meanBFScore metrics, the UNET's BIC_BIC was ranked the first thus demonstrating consistent superiority. In Figure 5, in terms of Accuracy, Dice score, and IoU, the SEGNET's BIC_BIC were not ranked the first, at the exception of meanBFScore. Finally, in Figure 6, in terms of ...

work page 2051
[5]

CONCLUSION Simultaneously avoiding the risk of undefined categorical labels and risk of exacerbating pixel -level annotation errors in data augmentation is a way to go in efforts to improve the quality of augmented data as demonstrated in this work. It was also demonstrated that only i mposing nearest -neighbor interpolation for handling categorical label...

work page
[6]

Data Augmentation in Classification and Segmentation: A Survey and New Strategies

Alomar K, Aysel HI, Cai X. Data Augmentation in Classification and Segmentation: A Survey and New Strategies. J Imaging. 2023 Feb 17;9(2):46. doi: 10.3390/jimaging9020046. PMID: 36826965; PMCID: PMC9966095

work page doi:10.3390/jimaging9020046 2023
[7]

Effects of Image Size on Deep Learning

Rukundo, O. Effects of Image Size on Deep Learning. Electronics 2023, 12, 985

work page 2023
[8]

Evaluation of Extra Pixel Interpolation Algorithm with Interpolation Mask Processing for Medical Image Segmentation with Deep Learning

Rukundo, O. Evaluation of Extra Pixel Interpolation Algorithm with Interpolation Mask Processing for Medical Image Segmentation with Deep Learning. SIViP (2024)

work page 2024
[9]

Rukundo, O., Non -extra Pixel Interpolation, International Journal of Image and Graphics, 20(4), 2050031, 14 pages, 2020:

work page 2020
[10]

Rukundo, O., Evaluation of Rounding Functions in Nearest Neighbour Interpolation, International Journal of Computational Methods, 10 pages, 2021

work page 2021
[11]

Molecular Biology of the Cell 25, 457–469

Hollandi R, Diósdi Á, Hollandi G, Moshkov N, Horváth P. AnnotatorJ: an ImageJ plugin to ease hand annotation of cellular compartments. Mol Biol Cell. 2020 Sep 15;31(20):2179 -2186. doi: 10.1091/mbc. E20 -02-

work page doi:10.1091/mbc 2020
[12]

PMID: 32697683; PMCID: PMC7550707

Epub 2020 Jul 22. PMID: 32697683; PMCID: PMC7550707

work page 2020
[13]

Rukundo, O., Normalized Weighting Schemes for Image Interpolation Algorithms, Applied Sciences, 13(3):1741, 16 pages, 2023:

work page 2023
[14]

13-22, 2022,

Rukundo, O., Schmidt, S., Stochastic Rounding for Image Interpolation and Scan Conversion, International Journal of Advanced Computer Science and Applications, 13(3), p. 13-22, 2022,

work page 2022
[15]

Kaggle Datasets < https://www.kaggle.com/datasets >, May 2023

work page 2023
[16]

Nikhil Tomar, Brain Tumor Segmentation, <https://www.kaggle.com/datasets/nikhilroxtomar/brain-tumor- segmentation>

work page
[17]

Debesh Jha, KvasirCapsule-SEG (capsule endoscopy dataset): https://www.kaggle.com/datasets/debeshjha1/kvasircapsuleseg

work page
[18]

U-Net: Convolutional networks for biomedical image segmentation

Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention —MICCAI 2015, Proceedings of the 18th International Conference, Munich, Germany, 5–9 October 2015

work page 2015
[19]

SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation,

V. Badrinarayanan, A. Kendall and R. Cipolla, "SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no.12, pp. 2481-2495, 1 Dec. 2017

work page 2017
[20]

Chen, LC., Zhu, Y., Papandreou, G., Schroff, F., Adam, H. (2018). Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds) Computer Vision – ECCV 2018. ECCV 2018. Lecture Notes in Computer Science, vol 11211. Springer, Cham

work page 2018
[21]

Convolutional Neural Networks for Automatic Detection of Intact Adenovirus from TEM Imaging with Debris, Broken and Artefacts Particles: ArXiv: 2310.19630

Rukundo, O., Behonova, A., et al. Convolutional Neural Networks for Automatic Detection of Intact Adenovirus from TEM Imaging with Debris, Broken and Artefacts Particles: ArXiv: 2310.19630

work page arXiv
[22]

The Use of Ranks to Avoid the Assumption of Normality Implicit in the Analysis of Variance

Friedman, M. The Use of Ranks to Avoid the Assumption of Normality Implicit in the Analysis of Variance. Journal of the American Statistical Association, 32(200), 675–701, 1937

work page 1937

[1] [1]

INTRODUCTION Data augmentation operation involve s many techniques that operate at the data level rather than at the convolutional neural network (CNN) architectural level [1]. In the context of semantic segmentation with deep learning , this operation aims at increasing both training sets images and corresponding masks to help the CNN architecture of int...

work page

[2] [2]

Datasets Three datasets were downloaded from Kaggle, a large platform hosting community -shared models, datasets, and codes [9]

MATERIALS AND METHODS A. Datasets Three datasets were downloaded from Kaggle, a large platform hosting community -shared models, datasets, and codes [9]. While selecting these datasets, the author focused on the diversity in digital imaging modalities and task of interest . The first dataset consists of fifty-five histology breast light- microscopy images...

work page

[3] [3]

Selected Evaluation Metrics To evaluate the quality of predictions made by CNN architectures under various interpolation settings , Accuracy (Eq

EXPERIMENTAL RESULTS A. Selected Evaluation Metrics To evaluate the quality of predictions made by CNN architectures under various interpolation settings , Accuracy (Eq. 4), IoU (Eq. 5), and meanBFScore (Eq. 6 and Eq. 7), as well as the Dice score (Eq. 8), were used. These metrics were chosen due to their well -established suitability for assessing segmen...

work page

[4] [4]

In Figure 5, in terms of Accuracy, Dice score, and IoU, the SEGNET's BIC_BIC were not ranked the first, at the exception of meanBFScore

DISCUSSIONS As can be seen in Figure 4, in terms of Accuracy, Dice score, IoU and meanBFScore metrics, the UNET's BIC_BIC was ranked the first thus demonstrating consistent superiority. In Figure 5, in terms of Accuracy, Dice score, and IoU, the SEGNET's BIC_BIC were not ranked the first, at the exception of meanBFScore. Finally, in Figure 6, in terms of ...

work page 2051

[5] [5]

CONCLUSION Simultaneously avoiding the risk of undefined categorical labels and risk of exacerbating pixel -level annotation errors in data augmentation is a way to go in efforts to improve the quality of augmented data as demonstrated in this work. It was also demonstrated that only i mposing nearest -neighbor interpolation for handling categorical label...

work page

[6] [6]

Data Augmentation in Classification and Segmentation: A Survey and New Strategies

Alomar K, Aysel HI, Cai X. Data Augmentation in Classification and Segmentation: A Survey and New Strategies. J Imaging. 2023 Feb 17;9(2):46. doi: 10.3390/jimaging9020046. PMID: 36826965; PMCID: PMC9966095

work page doi:10.3390/jimaging9020046 2023

[7] [7]

Effects of Image Size on Deep Learning

Rukundo, O. Effects of Image Size on Deep Learning. Electronics 2023, 12, 985

work page 2023

[8] [8]

Evaluation of Extra Pixel Interpolation Algorithm with Interpolation Mask Processing for Medical Image Segmentation with Deep Learning

Rukundo, O. Evaluation of Extra Pixel Interpolation Algorithm with Interpolation Mask Processing for Medical Image Segmentation with Deep Learning. SIViP (2024)

work page 2024

[9] [9]

Rukundo, O., Non -extra Pixel Interpolation, International Journal of Image and Graphics, 20(4), 2050031, 14 pages, 2020:

work page 2020

[10] [10]

Rukundo, O., Evaluation of Rounding Functions in Nearest Neighbour Interpolation, International Journal of Computational Methods, 10 pages, 2021

work page 2021

[11] [11]

Molecular Biology of the Cell 25, 457–469

Hollandi R, Diósdi Á, Hollandi G, Moshkov N, Horváth P. AnnotatorJ: an ImageJ plugin to ease hand annotation of cellular compartments. Mol Biol Cell. 2020 Sep 15;31(20):2179 -2186. doi: 10.1091/mbc. E20 -02-

work page doi:10.1091/mbc 2020

[12] [12]

PMID: 32697683; PMCID: PMC7550707

Epub 2020 Jul 22. PMID: 32697683; PMCID: PMC7550707

work page 2020

[13] [13]

Rukundo, O., Normalized Weighting Schemes for Image Interpolation Algorithms, Applied Sciences, 13(3):1741, 16 pages, 2023:

work page 2023

[14] [14]

13-22, 2022,

Rukundo, O., Schmidt, S., Stochastic Rounding for Image Interpolation and Scan Conversion, International Journal of Advanced Computer Science and Applications, 13(3), p. 13-22, 2022,

work page 2022

[15] [15]

Kaggle Datasets < https://www.kaggle.com/datasets >, May 2023

work page 2023

[16] [16]

Nikhil Tomar, Brain Tumor Segmentation, <https://www.kaggle.com/datasets/nikhilroxtomar/brain-tumor- segmentation>

work page

[17] [17]

Debesh Jha, KvasirCapsule-SEG (capsule endoscopy dataset): https://www.kaggle.com/datasets/debeshjha1/kvasircapsuleseg

work page

[18] [18]

U-Net: Convolutional networks for biomedical image segmentation

Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention —MICCAI 2015, Proceedings of the 18th International Conference, Munich, Germany, 5–9 October 2015

work page 2015

[19] [19]

SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation,

V. Badrinarayanan, A. Kendall and R. Cipolla, "SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no.12, pp. 2481-2495, 1 Dec. 2017

work page 2017

[20] [20]

Chen, LC., Zhu, Y., Papandreou, G., Schroff, F., Adam, H. (2018). Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds) Computer Vision – ECCV 2018. ECCV 2018. Lecture Notes in Computer Science, vol 11211. Springer, Cham

work page 2018

[21] [21]

Convolutional Neural Networks for Automatic Detection of Intact Adenovirus from TEM Imaging with Debris, Broken and Artefacts Particles: ArXiv: 2310.19630

Rukundo, O., Behonova, A., et al. Convolutional Neural Networks for Automatic Detection of Intact Adenovirus from TEM Imaging with Debris, Broken and Artefacts Particles: ArXiv: 2310.19630

work page arXiv

[22] [22]

The Use of Ranks to Avoid the Assumption of Normality Implicit in the Analysis of Variance

Friedman, M. The Use of Ranks to Avoid the Assumption of Normality Implicit in the Analysis of Variance. Journal of the American Statistical Association, 32(200), 675–701, 1937

work page 1937