Chaos-Enhanced Prototypical Networks for Few-Shot Medical Image Classification

Chinthakuntla Meghan Sai; Karthik Seemakurthy; Murarisetty V Sai Kartheek; Sita Devi Bharatula

arxiv: 2604.17300 · v1 · submitted 2026-04-19 · 📡 eess.IV · cs.AI· cs.CV

Chaos-Enhanced Prototypical Networks for Few-Shot Medical Image Classification

Chinthakuntla Meghan Sai , Murarisetty V Sai Kartheek , Sita Devi Bharatula , Karthik Seemakurthy This is my paper

Pith reviewed 2026-05-10 05:54 UTC · model grok-4.3

classification 📡 eess.IV cs.AIcs.CV

keywords few-shot learningprototypical networkschaos theorybrain tumor classificationmedical image classificationlogistic mapepisodic training

0 comments

The pith

Controlled chaotic perturbations during training help prototypical networks resist morphological noise in few-shot brain tumor scans.

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

Few-shot learning often fails on clinical scans because high intra-class variance and morphological noise scatter the learned prototypes. The authors integrate a logistic chaos map into a ResNet-18 backbone so that controlled random-like perturbations are added to the support-set features while the model learns. These deterministic yet ergodic injections act as low-cost stress tests that push the embedding space toward noise-invariant clusters. On a 4-way 5-shot brain tumor task the 15-percent injection level produced the highest accuracy, reaching 84.52 percent and beating a standard ProtoNet baseline. The approach therefore offers a simple regularization device that exploits the map's built-in mixing properties without extra parameters or heavy computation.

Core claim

By injecting controlled perturbations drawn from the logistic chaos map into the support features of each episode, the network learns embeddings whose class prototypes remain stable even when the input images contain the morphological variations typical of brain tumor MRI.

What carries the argument

Logistic chaos map injection module that adds deterministic yet ergodic perturbations to support-set feature vectors during episodic training.

If this is right

The same chaos injection schedule can be dropped into any metric-based few-shot learner that relies on episodic support sets.
Lower prototype variance should translate into more reliable decision boundaries when only five examples per class are available.
Because the perturbation is generated by a one-dimensional recurrence, the added compute is negligible compared with the backbone forward pass.

Where Pith is reading between the lines

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

If the map's ergodicity is the active ingredient, other low-dimensional chaotic maps with similar mixing properties might substitute with little loss of performance.
The reported accuracy gain could be re-interpreted as a form of deterministic data augmentation inside the feature space rather than image space.
Testing on multi-site or multi-scanner brain-tumor cohorts would reveal whether the noise-invariance generalizes beyond the single acquisition protocol used here.

Load-bearing premise

The logistic chaos map at the chosen 15 percent injection strength will consistently produce noise-invariant representations without creating new instabilities or needing per-dataset retuning.

What would settle it

A controlled experiment showing that the same 15 percent injection level either drops accuracy below the plain ProtoNet baseline or increases prototype dispersion on a different medical imaging dataset.

read the original abstract

The scarcity of labeled clinical data in oncology makes Few-Shot Learning (FSL) a critical framework for Computer Aided Diagnostics, but we observed that standard Prototypical Networks often struggle with the "prototype instability" caused by morphological noise and high intra-class variance in brain tumor scans. Our work attempts to minimize this by integrating a non-linear Logistic Chaos Module into a fine-tuned ResNet-18 backbone creating the Chaos-Enhanced ProtoNet(CE-ProtoNet). Using the deterministic ergodicity of the logistic chaos map we inject controlled perturbations into support features during episodic training-essentially for "stress testing" the embedding space. This process makes the model to converge on noise-invariant representations without increasing computational overhead. Testing this on a 4-way 5-shot brain tumor classification task, we found that a 15% chaotic injection level worked efficiently to stabilize high-dimensional clusters and reduce class dispersion. Our method achieved a peak test accuracy of 84.52%, outperforming standard ProtoNet. Our results suggest the idea of using chaotic perturbation as an efficient, low-overhead regularization tool, for the data-scarce regimes.

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 adds logistic chaos perturbations to prototypical networks for 4-way 5-shot brain tumor classification and reports 84.52% accuracy, but the results do not show that the chaos map itself drives the gain over ordinary regularization.

read the letter

The paper integrates a logistic chaos module into a fine-tuned ResNet-18 prototypical network to inject controlled perturbations into support features during episodic training. The goal is to reduce prototype instability from morphological variance in brain tumor scans. They test on a 4-way 5-shot task and report a peak accuracy of 84.52%, beating the standard ProtoNet baseline. The approach keeps overhead low and targets a genuine pain point in medical few-shot learning where labeled data is scarce. That combination of an existing chaos map with this specific backbone and task is the concrete new element. It is a straightforward extension rather than a conceptual shift, but it is applied cleanly to the medical setting. The motivation is clear and the implementation sounds simple enough to reproduce from the description. The main weakness is the evaluation. Only a single peak accuracy number appears, with no error bars, no multiple random seeds, and no statistical tests. The 15% injection level is described as the one that worked efficiently, which points to post-hoc selection. There are no ablations against matched non-chaotic noise such as Gaussian perturbations or dropout, and no direct measures of class dispersion or noise invariance are given. Without those, it is impossible to tell whether the deterministic ergodicity of the logistic map is load-bearing or whether any perturbation of similar strength would produce the same effect. The central claim about noise-invariant representations therefore rests on an untested assumption. This work is aimed at applied researchers who already use prototypical networks in medical imaging and are looking for lightweight regularization ideas. A reader could pull the module, plug it into their own pipeline, and check the effect on their data. It is not ready for strong claims yet, but the idea is concrete and the problem is relevant, so it deserves a serious referee who can ask for the missing controls and comparisons. I would send it to review.

Referee Report

4 major / 2 minor

Summary. The paper proposes Chaos-Enhanced Prototypical Networks (CE-ProtoNet) that augment a fine-tuned ResNet-18 backbone with a Logistic Chaos Module. During episodic training, deterministic perturbations drawn from the logistic map are injected into support-set features at a 15% level; the authors claim this stabilizes high-dimensional prototypes against morphological noise in brain-tumor images, yielding noise-invariant embeddings and a peak 4-way 5-shot test accuracy of 84.52% that exceeds standard ProtoNet.

Significance. If the reported gain is reproducible and attributable to the ergodic properties of the logistic map rather than generic regularization, the method would supply a low-overhead, parameter-light regularization strategy for few-shot medical imaging. The absence of ablations, error bars, or invariance metrics currently prevents any such assessment.

major comments (4)

[Abstract] Abstract: the sole quantitative result is a single peak accuracy of 84.52% on one 4-way 5-shot split; no mean accuracy, standard deviation, number of random seeds, or statistical comparison to the ProtoNet baseline is supplied, rendering the central performance claim unverifiable.
[Abstract] Abstract: the 15% injection level is stated to have 'worked efficiently' without any sensitivity sweep, ablation table, or justification that other percentages (or non-chaotic noise of matched magnitude) fail to produce comparable gains.
[Abstract] Abstract: no quantitative proxy for the claimed 'stabilization of high-dimensional clusters' or 'reduction in class dispersion' (e.g., intra-class variance, prototype drift, or embedding invariance metrics) is reported, so the mechanistic explanation remains unsupported.
[Abstract] Abstract: the paper does not compare logistic-map injection against matched-strength alternatives (Gaussian noise, dropout, or random affine perturbations), leaving open whether the deterministic ergodicity of the map is load-bearing or whether any structured perturbation suffices.

minor comments (2)

[Abstract] Abstract contains minor grammatical issues ('makes the model to converge', 'Our results suggest the idea of using').
[Abstract] The abstract omits the dataset name, total number of images, train/val/test split sizes, and the precise definition of the 4-way 5-shot episode construction.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment point by point below and will revise the manuscript to incorporate additional analyses and clarifications.

read point-by-point responses

Referee: [Abstract] Abstract: the sole quantitative result is a single peak accuracy of 84.52% on one 4-way 5-shot split; no mean accuracy, standard deviation, number of random seeds, or statistical comparison to the ProtoNet baseline is supplied, rendering the central performance claim unverifiable.

Authors: We acknowledge that the abstract presents only the peak accuracy for brevity. We will revise the abstract to report mean accuracy and standard deviation over multiple random seeds, along with a statistical comparison to the ProtoNet baseline. revision: yes
Referee: [Abstract] Abstract: the 15% injection level is stated to have 'worked efficiently' without any sensitivity sweep, ablation table, or justification that other percentages (or non-chaotic noise of matched magnitude) fail to produce comparable gains.

Authors: We agree that a sensitivity analysis and comparison to non-chaotic alternatives would strengthen the justification. We will add an ablation table varying the injection level and comparing to non-chaotic noise of matched magnitude. revision: yes
Referee: [Abstract] Abstract: no quantitative proxy for the claimed 'stabilization of high-dimensional clusters' or 'reduction in class dispersion' (e.g., intra-class variance, prototype drift, or embedding invariance metrics) is reported, so the mechanistic explanation remains unsupported.

Authors: We will include quantitative metrics such as intra-class variance, prototype drift, and embedding invariance measures in the results section to support the claims of stabilization. revision: yes
Referee: [Abstract] Abstract: the paper does not compare logistic-map injection against matched-strength alternatives (Gaussian noise, dropout, or random affine perturbations), leaving open whether the deterministic ergodicity of the map is load-bearing or whether any structured perturbation suffices.

Authors: We will add comparisons of logistic-map injection against Gaussian noise, dropout, and random affine perturbations at matched strengths to demonstrate the role of the chaotic properties. revision: yes

Circularity Check

0 steps flagged

No circularity detected in empirical method

full rationale

The paper presents a purely empirical contribution: integrating a logistic chaos module into a ResNet-18 backbone for prototypical networks and reporting test accuracy on a 4-way 5-shot brain tumor task. No equations, derivations, or parameter-fitting steps appear in the abstract or described text. The central result (84.52% peak accuracy at 15% injection) is an observed experimental outcome rather than a quantity derived from or defined in terms of itself. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing elements. The work is self-contained as standard empirical ML research with no reduction of claims to fitted inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of the chaos injection, with the 15% level acting as a tuned hyperparameter and the ergodicity of the logistic map treated as a domain assumption that delivers controlled perturbations.

free parameters (1)

chaotic injection level = 15%
Described as the level that 'worked efficiently' to stabilize clusters, indicating it was selected based on experimental outcomes rather than a priori.

axioms (1)

domain assumption The logistic map provides deterministic ergodic perturbations suitable for stress-testing embeddings to achieve noise-invariant representations without added computational cost.
Invoked when describing the Chaos Module and its integration during episodic training.

pith-pipeline@v0.9.0 · 5519 in / 1537 out tokens · 70569 ms · 2026-05-10T05:54:01.385327+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[1]

Chowdhury, R.R., Bathula, D.R.: Refining Prototypes with Selective Feature Maps: A Lightweight Approach to Few-Shot Medical Image Classification. Int. J. Comput. Appl.975(1), 8887 (2025)

work page 2025
[2]

Quan, H., Li, X., Hu, D., Nan, T., Cui, X.: Dual-channel prototype network for few-shot pathology image classification. IEEE J. Biomed. Health Inform.28(7), 4132–4144 (2024)

work page 2024
[3]

IEEE/ACM Trans

Lu, L., Cui, X., Tan, Z., Wu, Y.: MedOptNet: Meta-learning framework for few- shot medical image classification. IEEE/ACM Trans. Comput. Biol. Bioinform. 21(4), 725–736 (2023)

work page 2023
[4]

Jiang, H., Gao, M., Li, H., Jin, R., Miao, H., Liu, J.: Multi-learner based deep meta-learning for few-shot medical image classification. IEEE J. Biomed. Health Inform.27(1), 17–28 (2022)

work page 2022
[5]

Automatika66(2), 173–187 (2025)

Abirami, R., Malathy, C.: Secured DICOM medical image transition with opti- mized chaos method for encryption and customized deep learning model for watermarking. Automatika66(2), 173–187 (2025)

work page 2025
[6]

Mathematics11(4), 1032 (2023)

Dahou, A., Chelloug, S.A., Alduailij, M., Elaziz, M.A.: Improved feature selection based on chaos game optimization for social internet of things with a novel deep learning model. Mathematics11(4), 1032 (2023)

work page 2023
[7]

Mabrouk, A., Dahou, A., Elaziz, M.A., D´ ıaz Redondo, R.P., Kayed, M.: Medical image classification using transfer learning and chaos game optimization on the 16 internet of medical things. Comput. Intell. Neurosci.2022, 9112634 (2022)

work page 2022
[8]

Abdel-Salam, M., Hu, G., C ¸ elik, E., Gharehchopogh, F.S., El-Hasnony, I.M.: Chaotic RIME optimization algorithm with adaptive mutualism for feature selection problems. Comput. Biol. Med.179, 108803 (2024)

work page 2024
[9]

IEEE Access8, 186638–186652 (2020)

Elgamal, Z.M., Yasin, N.B., Tubishat, M., Alswaitti, M., Mirjalili, S.: An improved harris hawks optimization algorithm with simulated annealing for feature selection in the medical field. IEEE Access8, 186638–186652 (2020)

work page 2020
[10]

Image Vis

Ouahab, A., Ahmed, O.B.: ProtoMed: Prototypical networks with auxiliary reg- ularization for few-shot medical image classification. Image Vis. Comput.154, 105337 (2025)

work page 2025
[11]

Cheng, Z., Zhao, J., Deng, J., Zhang, H.: Few-Shot Medical Image Segmentation with High-Confidence Prior Mask. IEEE J. Biomed. Health Inform. (2025). https: //doi.org/10.1109/JBHI.2025.3371234

work page doi:10.1109/jbhi.2025.3371234 2025
[12]

IEEE Trans

Shen, Y., Fan, W., Wang, C., Liu, W., Wang, W., Zhang, Q., Zhou, D.: Dual- guided prototype alignment network for few-shot medical image segmentation. IEEE Trans. Instrum. Meas.73, 1–13 (2024)

work page 2024
[13]

In: MICCAI Workshop on Data Engineering in Medical Imaging, pp

Uddin, I.I., Wang, L., Santosh, K.C.: Expert-guided explainable few-shot learn- ing for medical image diagnosis. In: MICCAI Workshop on Data Engineering in Medical Imaging, pp. 95–104. Springer, Cham (2025)

work page 2025
[14]

PLoS ONE18(6), e0286915 (2023)

Liu, T., Ke, Z., Li, Y., Silamu, W.: Knowledge-enhanced prototypical network with class cluster loss for few-shot relation classification. PLoS ONE18(6), e0286915 (2023)

work page 2023
[15]

Diagnostics15(18), 2404 (2025)

Qui˜ nonez-Baca, L.C., Ramirez-Alonso, G., Gaxiola, F., Manzo-Martinez, A., Cornejo, R., Lopez-Flores, D.R.: A Comparative Evaluation of Meta-Learning Models for Few-Shot Chest X-Ray Disease Classification. Diagnostics15(18), 2404 (2025)

work page 2025
[16]

npj Digit

Yu, Z., Nguyen, T.D., Ju, L., Gal, Y., Sashindranath, M., Bonnington, P., Zhang, L., Mar, V., Ge, Z.: Hierarchical skin lesion image classification with prototypical decision tree. npj Digit. Med.8(1), 26 (2025)

work page 2025
[17]

PLoS ONE19(5), e0298227 (2024)

Awudong, B., Li, Q., Liang, Z., Tian, L., Yan, J.: Attentional adversarial training for few-shot medical image segmentation without annotations. PLoS ONE19(5), e0298227 (2024)

work page 2024
[18]

Hossain, M.A., Sa˘ g, T.: Chaotic Levy-Flight-Driven Siberian Tiger Optimization for Enhanced Data Clustering. Cybern. Syst.57(1), 44–83 (2026) 17

work page 2026
[19]

Ali, M.D., Iqbal, M.A., Lee, S., Duan, X., Kim, S.K.: Explainable AI Based Multi Class Skin Cancer Detection Enhanced by Meta Learning with Generative DDPM Data Augmentation. Appl. Sci.15(21), 11689 (2025)

work page 2025
[20]

Neurocomputing130774 (2025)

Lim, J.Y., Lim, K.M., Lee, C.P., Tan, Y.X.: A review of few-shot image clas- sification: Approaches, datasets and research trends. Neurocomputing130774 (2025)

work page 2025
[21]

In: International Conference on Learning Representations (ICLR) (2019)

Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (ICLR) (2019)

work page 2019
[22]

Advances in Neural Information Processing Systems (NeurIPS)31(2018)

Oreshkin, B., L´ opez, P.R., Lacoste, A.: TADAM: Task dependent adaptive met- ric for improved few-shot learning. Advances in Neural Information Processing Systems (NeurIPS)31(2018)

work page 2018
[23]

Kaggle (2025)

Bhuvaji, S., Kadam, A., Bhumkar, P., Dedge, S., Kanchan, S.: Brain Tumor Clas- sification (MRI) [Data set]. Kaggle (2025). https://doi.org/10.34740/KAGGLE/ DSV/12745533 18

work page doi:10.34740/kaggle/ 2025

[1] [1]

Chowdhury, R.R., Bathula, D.R.: Refining Prototypes with Selective Feature Maps: A Lightweight Approach to Few-Shot Medical Image Classification. Int. J. Comput. Appl.975(1), 8887 (2025)

work page 2025

[2] [2]

Quan, H., Li, X., Hu, D., Nan, T., Cui, X.: Dual-channel prototype network for few-shot pathology image classification. IEEE J. Biomed. Health Inform.28(7), 4132–4144 (2024)

work page 2024

[3] [3]

IEEE/ACM Trans

Lu, L., Cui, X., Tan, Z., Wu, Y.: MedOptNet: Meta-learning framework for few- shot medical image classification. IEEE/ACM Trans. Comput. Biol. Bioinform. 21(4), 725–736 (2023)

work page 2023

[4] [4]

Jiang, H., Gao, M., Li, H., Jin, R., Miao, H., Liu, J.: Multi-learner based deep meta-learning for few-shot medical image classification. IEEE J. Biomed. Health Inform.27(1), 17–28 (2022)

work page 2022

[5] [5]

Automatika66(2), 173–187 (2025)

Abirami, R., Malathy, C.: Secured DICOM medical image transition with opti- mized chaos method for encryption and customized deep learning model for watermarking. Automatika66(2), 173–187 (2025)

work page 2025

[6] [6]

Mathematics11(4), 1032 (2023)

Dahou, A., Chelloug, S.A., Alduailij, M., Elaziz, M.A.: Improved feature selection based on chaos game optimization for social internet of things with a novel deep learning model. Mathematics11(4), 1032 (2023)

work page 2023

[7] [7]

Mabrouk, A., Dahou, A., Elaziz, M.A., D´ ıaz Redondo, R.P., Kayed, M.: Medical image classification using transfer learning and chaos game optimization on the 16 internet of medical things. Comput. Intell. Neurosci.2022, 9112634 (2022)

work page 2022

[8] [8]

Abdel-Salam, M., Hu, G., C ¸ elik, E., Gharehchopogh, F.S., El-Hasnony, I.M.: Chaotic RIME optimization algorithm with adaptive mutualism for feature selection problems. Comput. Biol. Med.179, 108803 (2024)

work page 2024

[9] [9]

IEEE Access8, 186638–186652 (2020)

Elgamal, Z.M., Yasin, N.B., Tubishat, M., Alswaitti, M., Mirjalili, S.: An improved harris hawks optimization algorithm with simulated annealing for feature selection in the medical field. IEEE Access8, 186638–186652 (2020)

work page 2020

[10] [10]

Image Vis

Ouahab, A., Ahmed, O.B.: ProtoMed: Prototypical networks with auxiliary reg- ularization for few-shot medical image classification. Image Vis. Comput.154, 105337 (2025)

work page 2025

[11] [11]

Cheng, Z., Zhao, J., Deng, J., Zhang, H.: Few-Shot Medical Image Segmentation with High-Confidence Prior Mask. IEEE J. Biomed. Health Inform. (2025). https: //doi.org/10.1109/JBHI.2025.3371234

work page doi:10.1109/jbhi.2025.3371234 2025

[12] [12]

IEEE Trans

Shen, Y., Fan, W., Wang, C., Liu, W., Wang, W., Zhang, Q., Zhou, D.: Dual- guided prototype alignment network for few-shot medical image segmentation. IEEE Trans. Instrum. Meas.73, 1–13 (2024)

work page 2024

[13] [13]

In: MICCAI Workshop on Data Engineering in Medical Imaging, pp

Uddin, I.I., Wang, L., Santosh, K.C.: Expert-guided explainable few-shot learn- ing for medical image diagnosis. In: MICCAI Workshop on Data Engineering in Medical Imaging, pp. 95–104. Springer, Cham (2025)

work page 2025

[14] [14]

PLoS ONE18(6), e0286915 (2023)

Liu, T., Ke, Z., Li, Y., Silamu, W.: Knowledge-enhanced prototypical network with class cluster loss for few-shot relation classification. PLoS ONE18(6), e0286915 (2023)

work page 2023

[15] [15]

Diagnostics15(18), 2404 (2025)

Qui˜ nonez-Baca, L.C., Ramirez-Alonso, G., Gaxiola, F., Manzo-Martinez, A., Cornejo, R., Lopez-Flores, D.R.: A Comparative Evaluation of Meta-Learning Models for Few-Shot Chest X-Ray Disease Classification. Diagnostics15(18), 2404 (2025)

work page 2025

[16] [16]

npj Digit

Yu, Z., Nguyen, T.D., Ju, L., Gal, Y., Sashindranath, M., Bonnington, P., Zhang, L., Mar, V., Ge, Z.: Hierarchical skin lesion image classification with prototypical decision tree. npj Digit. Med.8(1), 26 (2025)

work page 2025

[17] [17]

PLoS ONE19(5), e0298227 (2024)

Awudong, B., Li, Q., Liang, Z., Tian, L., Yan, J.: Attentional adversarial training for few-shot medical image segmentation without annotations. PLoS ONE19(5), e0298227 (2024)

work page 2024

[18] [18]

Hossain, M.A., Sa˘ g, T.: Chaotic Levy-Flight-Driven Siberian Tiger Optimization for Enhanced Data Clustering. Cybern. Syst.57(1), 44–83 (2026) 17

work page 2026

[19] [19]

Ali, M.D., Iqbal, M.A., Lee, S., Duan, X., Kim, S.K.: Explainable AI Based Multi Class Skin Cancer Detection Enhanced by Meta Learning with Generative DDPM Data Augmentation. Appl. Sci.15(21), 11689 (2025)

work page 2025

[20] [20]

Neurocomputing130774 (2025)

Lim, J.Y., Lim, K.M., Lee, C.P., Tan, Y.X.: A review of few-shot image clas- sification: Approaches, datasets and research trends. Neurocomputing130774 (2025)

work page 2025

[21] [21]

In: International Conference on Learning Representations (ICLR) (2019)

Chen, W.Y., Liu, Y.C., Kira, Z., Wang, Y.C.F., Huang, J.B.: A closer look at few-shot classification. In: International Conference on Learning Representations (ICLR) (2019)

work page 2019

[22] [22]

Advances in Neural Information Processing Systems (NeurIPS)31(2018)

Oreshkin, B., L´ opez, P.R., Lacoste, A.: TADAM: Task dependent adaptive met- ric for improved few-shot learning. Advances in Neural Information Processing Systems (NeurIPS)31(2018)

work page 2018

[23] [23]

Kaggle (2025)

Bhuvaji, S., Kadam, A., Bhumkar, P., Dedge, S., Kanchan, S.: Brain Tumor Clas- sification (MRI) [Data set]. Kaggle (2025). https://doi.org/10.34740/KAGGLE/ DSV/12745533 18

work page doi:10.34740/kaggle/ 2025