List of Papers Browse by Subject Areas Author List
Abstract
Noisy labels in high-dimensional, and multiclass medical im-
age datasets pose a significant challenge for machine learning models.
While hybrid quantum-classical architectures, such as quantum neural
networks (QNNs), have shown promise in medical imaging, their ro-
bustness under noisy label conditions remains largely unexplored. To
address this gap, we propose a Noise-aware Quantum Neural Network
(NQNN), integrating Fourier Attenuation, Reweight Estimation, and
Adaptive Pooling to enhance feature extraction and classification ro-
bustness. Fourier Attenuation filters high-frequency noise, Reweight Es-
timation prioritizes cleaner labels based on uncertainty, and Adaptive
Pooling dynamically refines feature aggregation. We evaluate NQNN on
six benchmark medical datasets (PathMNIST, BloodMNIST, OrganAM-
NIST, OrganCMNIST, OCTMNIST, and DermaMNIST) across noise ra-
tios (10%, 30%, and 50%) and classification configurations (binary, four-
class, and full multiclass). Comparative benchmarks against five QNN-
based and two deep-learning baselines demonstrate NQNN’s superior
performance, such that achieving 80.25% accuracy on organCMNIST at
10% noise and maintaining strong performance at higher noise ratios.
Our ablation studies validate the effectiveness of each noise-handling
mechanism, highlighting their complementary contributions to noise ro-
bustness. By bridging quantum advancements with real-world medical di-
agnostics, NQNN establishes a new benchmark for noise-resilient medical
image classification, offering a scalable and adaptive quantum-classical
learning framework.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0804_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://drive.google.com/drive/folders/1zi3cZn4COkLhgNIqfJHtsApwdLUaFzqW?usp=drive_link
Link to the Dataset(s)
N/A
BibTex
@InProceedings{RahMaq_NQNN_MICCAI2025,
author = { Rahman, Maqsudur and Zhuang, Jun},
title = { { NQNN: Noise-aware Quantum Neural Networks for Medical Image Classification } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15972},
month = {September},
page = {434 -- 444}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a Noise-aware Quantum Neural Network (NQNN) framework to address the problem of label noise in medical image classification. The proposed method is evaluated against several baseline approaches across six medical image classification datasets with synthetic noisy label noise. Ablation studies are conducted to assess the contributions of key components of the framework.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper tackles the problem of learning with label noise, which is relatively underexplored in the context of quantum neural networks for medical image classification.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The motivation for proposing the NQNN framework is not clearly articulated. While the introduction highlights that label noise is an underexplored issue in QNNs, the paper does not sufficiently review existing label-noise learning methods in the classical setting, nor does it analyze their limitations when applied to QNNs. Without such a discussion, it is difficult to understand the necessity and advantage of the proposed framework.
- The methodological description lacks precision and clarity. Most of the symbols used in the equations are not defined, making it difficult to follow the implementation details. For example, how is label uncertainty estimated via disagreement in Eq. 1? What exactly do the terms “disagreement” and “single-target uncertainty” in Eq. 2 mean? What does p_c represent? This lack of explanation persists throughout Equations (3) to (7), which significantly hinders reproducibility and understanding.
- Although experiments were conducted on six synthetic noisy datasets, the absence of evaluations on real-world noisy datasets—such as LIDC-IDRI, CheXpert, or Kaggle DR—is a notable limitation. Evaluating the method under real-world noise conditions is essential to demonstrate its practical robustness. Moreover, the type and configuration of synthetic noise used in the experiments are not described. Is the noise symmetric, asymmetric, or instance-dependent? Such details are critical for interpreting the results.
- The selection of competing methods appears unfair. The first five methods are not specifically designed to handle label noise, and the two noise-robust methods included were published in 2021 and 2022. Recent state-of-the-art methods in learning with noisy labels are not considered, weakening the comparative claims. Additionally, comparisons on clean datasets (Table 1) seem unnecessary. A more in-depth analysis of NQNN’s robustness on real-world noisy data would be more informative and aligned with the paper’s main focus.
- The review of label-noise-robust learning methods in Section 4 is incomplete and loosely structured. Recent advances in noisy sample detection, noise transition matrix estimation, and noise-robust loss functions are largely omitted. Most cited works are from before 2022, which does not reflect the current state of the field.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(1) Strong Reject — must be rejected due to major flaws
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The motivation and detailed description of the proposed method are very unclear, the comparison methods are outdated, and no comparative experiments have been conducted on real-world noisy datasets.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Motivation is unclear. (1) What concrete advantages do QNNs offer for medical image classification? The claim that QNNs “show promise” lacks specific evidence. If QNNs struggle with high-resolution images—as acknowledged—how are they suitable for medical imaging, where fine details are crucial? Without clear benefits, the motivation for this study is weak. (2) Assuming QNNs are appropriate for medical image classification, it is valuable to explore how state-of-the-art noise-robust techniques can be adapted to them. However, the methods used in this work are outdated, and no specific adaptation challenges are discussed. If adaptation is trivial, the technical contribution is unclear; if it is non-trivial, the paper fails to articulate the core issues and solutions.
Experimental scope is limited. The paper only evaluates on symmetric noise. Real-world noise is typically instance-dependent, and most works benchmark across synthetic (symmetric, asymmetric, instance-dependent) and real-world noisy datasets. This narrow evaluation severely limits the significance of the results.
Review #2
- Please describe the contribution of the paper
The paper explores the robustness of quantum neural networks, in particular their NQNN architecture, to address the problem of noisy/incorrect labels in medical image datasets. It’s an important issue in machine learning, specifically medical imaging, affecting model performance. NQNN is a quantum-classical architecture designed to be robust integrating Fourier attention, Reweight estimation and adaptive pooling. Results on six MedMNIST datasets with varying simulated symmetric noise levels demonstrate NQNN’s superior performance over baselines (e.g., achieving 80.25% accuracy on OrganCMNIST at 10% noise) and maintaining good performance at higher noise ratios, also compared against 5 QNNs and 2 classical models.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(1) novelty in addressing the important but under-explored problem of QNN robustness to label noise within the relevant domain of medical imaging. (2) Considering quantum neural networks for medical imaging, particularly combining multiple noise-handling mechanisms. (3) Evaluation across various datasets, noise levels, and task complexities, supported by ablation studies, showing the relevance of all three mechanisms. (4) NQNN shows strong comparative performance, particularly against other QNNs, presented in the paper.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(1) The findings are based on classical simulations (understandably, of course, due to the lack of an actual quantum computer). Performance on real quantum hardware may differ significantly to the simulations. So it mostly shows potential future use.
(2) While it shows future potential, at the moment it is not clear if this will scale to larger datasets and in particular higher-resolution samples (beyond 28x28 MedMNist samples), relevant for clinical images. Of course simulation here is not feasible. Maybe discuss potential or optimisations for near-term devices.
(3) The scope of noise is limited and not fully explored, relying on synthetic noise with little detail on how it has been created (uniform, symmetric noise on the labels?). It is not clear how this relates to real-world noise patterns. Pleas specify more clearly how noise has been added.
(4) The chosen classical baselines might not represent the contemporary state-of-the-art (e.g., more recent methods leveraging Vision Language Models (e.g. DEFT/NoiseGPT) or correction techniques (e.g. DULC, TMLC-Net), even if difficult to compare on the particular dataset (I’m not aware of published results for these datasets), and the limitations of the quantum simulation. This potentially weakens the comparison; some justification for the choice of classical models is required.
(5) Inherent challenges in training VQCs (e.g. barren plateaus) were not discussed. It is not clear how optimised to the datasets presented the other quantum architectures are - the reported results appear somewhat low. A deeper exploration of the performance variations may be interesting.
(6) Given the MICCAI audience, some more details, if space permits, on the quantum mechanisms would be useful, potentially also with a discussion of why the improved performance may have been achieved by the three mechanisms introduced.
(7) The paper does not discuss the computational cost associated with classically simulating quantum circuits (minor, but relevant considering the uncertainties around quantum hardware).
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is an exploration of the future potential of quantum machine learning in medicine and shows some good results. More information is needed, in particular on the models and how the evaluation holds up (points 4/5 above) and the label noise (point 3).
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The paper presents an innovative approach to a significant problem in medical imaging, learning from noisy labels, by leveraging QNNs.
The rebuttal successfully addresses several critical points (clarified the type of noise used; provided reasonable justification for the choice of datasets; addressing concerns regarding dataset scale; explanation for baseline selection is largely acceptable, especially for QNN comparisons and the focus on label noise; simulation and NISQ-era considerations have been addressed; the authors have committed to improving methodological clarity and releasing code).
The somewhat remaining concern is the clarity of the methodology. While the authors promise to fix this, the decision must weigh the submitted version. However, the novelty of integrating specific noise-handling mechanisms within a VQC, combined with promising results on benchmark datasets (even if simulated and low-resolution), makes this a valuable contribution for forward-looking solutions. Limitations regarding immediate clinical scalability are inherent in most current quantum machine learning research and are acknowledged by the authors.
Assuming the authors address the methodological clarity as promised, I’m happy to recommend acceptance - the rebuttal has sufficiently alleviated the concerns.
Review #3
- Please describe the contribution of the paper
The authors formulate a mechanism for dealing with labeling noise in training quantum neural networks. They combine three noise-handling techniques: Fourier attenuation, reweight estimation, and adaptive pooling. This would be mildly interesting on its own, but is made significantly more so by the requirement that it needs a quantum computer to operate.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper utilizes an emerging computing technology to solve a real-world problem with medical imaging training data. The authors provide a comprehensive evaluation and compare against other non-quantum approaches and show several benchmarks. The technique outperforms all QNN baselines and in some cases surpasses classical methods.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
They do not provide a good discussion of the quantum hardware requirements and limitations. The solutions they show utilize toy images of 28x28, but scaling will be a problem since it will require significantly more qubits. It would also be nice to see the technique validated on real quantum hardware. They do not go into detail on what hardware they are using other than it is an Ubuntu system. Does it run on GPU? Some quantum simulator? More detail is needed.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While I hope to see a more detailed discussion of requirements for quantum hardware to scale this technique to full sized images, I think that the theory is sound and well demonstrated.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and area chairs for their constructive feedback and use this opportunity to clarify our contributions to NQNN by addressing key concerns and refining the presentation, with all clarifications incorporated and full code to be released upon acceptance. Motivation and Contribution(R2-C1, C4) Our goal is to design a quantum-classical model that is robust to label noise in multiclass medical image classification, especially in scenarios with misannotated labels (sec 1, p-1). While classical noisy-image learning is well studied, QNNs have not been adapted for such settings. We propose NQNN, a hybrid model embedding three noise-resilient layers directly into the VQC. Since prior QNNs are designed for clean-label classification, we benchmarked NQNN on both clean and noisy data to demonstrate its added resilience without degrading performance under standard conditions. Methodology, Equation Clarity, and Reproducibility(R2-C2, R1-C6) All variables in Equations (1)–(7), including Pc, Ud, and Us, are defined in our implementation code. These terms govern dynamic sample weighting in our circuit. Due to space constraints, some definitions were abbreviated, but we will revise the final manuscript to clearly describe all components without altering any results. As stated in Sec. 3, p-4 we will release the full codebase upon acceptance, including annotated implementations, preprocessing steps, and model configurations for reproducibility. Dataset and Label Noise Strategy(R1-C2&3, R2-C3) We selected six datasets from MedMNISTv2, which offer diverse modalities and up to 11-class tasks while remaining compatible with current quantum simulation limits. Unlike high-resolution datasets such as CheXpert or LIDC-IDRI that involve image-content noise, MedMNISTv2’s 28×28 structured inputs align with the limitations of current quantum circuit simulation, especially when integrating our three noise-handling layers to study label-level noise by introducing symmetric class-flipping at certain noise rates. These settings allow us to stress the model’s ability to generalize under noisy labels in realistic yet controlled conditions, aligned with our architectural scope. Hardware, Scalability, Simulation Cost(R1-C1,5,7; R3) All experiments were conducted using the Pennylane default.qubit simulator with GPU acceleration. Classical simulation is computationally expensive but essential for prototyping QNN architectures. To ensure hardware feasibility, we used shallow circuits (5 layers, 8 qubits), which avoid barren plateaus and remain compatible with NISQ-era constraints. This setup provides a practical foundation for evaluating noise-resilient layers within current quantum limits and paves the way for future scaling to higher-resolution medical imaging. Baseline Selection(R1-C4, R2-C4) Our objective is to assess how existing QNNs perform under noisy labels and to demonstrate the effectiveness of integrating noise-resilient mechanisms within VQC. As no prior work explicitly addresses noisy-label robustness in QNNs particularly in multiclass medical imaging, we selected five representative QNN baselines spanning diverse architectures. Although developed for clean-label scenarios, comparing them under both clean and noisy conditions allowed us to reveal significance of our proposed layers. We selected DUE-Net and RWNet for their focus on noisy-label learning and reproducibility, while newer methods like DULC and NoiseGPT target image-content noise and high-resolution tasks, making them less applicable to our label-noise-focused, low-dimensional quantum framework. Related Work Coverage(R2-C5) Section 4 was intended as a focused contextual review. While not exhaustive, we included seven recent 2024 works alongside foundational studies most relevant to our goal. Page limits constrained further expansion in both paper and reference, but our contribution lies in adapting label-noise robustness to the quantum domain, an area with little prior work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A