acceptodds
Withdrawn from ICLR 2027

SLT: Robust Quantum Neural Networks for Noisy-Label Medical Image Classification via Supermartingale-based Label Transition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Noisy-label learning in small-scale medical image classification remains challenging, as limited data can amplify annotation errors. While quantum neural networks (QNNs) have recently been explored as compact models for small-data regimes, their behavior under label corruption remains insufficiently understood. A key obstacle is QNNs' intrinsic "natural smoothness", which may regularize training but also obscure high-confidence samples needed for noise-transition estimation. We propose Supermartingale-based Label Transition (SLT), an anchor-free loss correction framework for robust QNN-based medical image classification under noisy labels. SLT models entropy reduction in predictive distributions as a supermartingale and uses its monotonic behavior to identify stable transition-matrix refinement steps. This enables dynamic transition updates while reducing noise-driven oscillations during QNN training. We further provide a convergence analysis showing that the proposed transition-refinement process reaches a steady state. Experiments on multiple public small-scale medical image datasets demonstrate that SLT improves QNN-based noisy-label classification and outperforms classic noisy-label learning baselines under synthetic and real-world label noise.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
medical image analysis, noisy label learning, quantum machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_c5YROpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies noisy-label medical image classification with hybrid quantum neural networks. It proposes SLT, which uses the running minimum of the mean predictive entropy to decide when to update a label-transition matrix. At an update step, the method estimates the co-occurrence of predicted classes and observed noisy labels, then combines this estimate with the previous matrix through a moving average. The resulting matrix is used for forward loss correction. The authors characterize the running entropy minimum as a supermartingale and use this to argue that the transition matrix converges. Experiments cover five MedMNIST datasets with synthetic label corruption, with additional evaluations under instance-dependent noise and on CheXpert. The paper also compares quantum and classical backbones and studies temperature scaling, circuit size, and update schedules.

Critical strengths
  1. The paper addresses a useful question: whether label-noise correction can improve small hybrid quantum classifiers. The proposed update is simple, using prediction-label counts and a moving average rather than introducing another trainable noise-estimation network.

  2. The experimental evaluation includes several medical imaging modalities, multiple corruption settings, and five independent runs with reported standard deviations. The main loss-correction baselines use the same QNN backbone, which helps assess the effect of the correction method within this model family. The reported gains are generally consistent across the included datasets.

Critical weaknesses
  1. Theorem 2 is not established by the proof in Appendix A. The proof argues that because the decreases in the running entropy minimum tend to zero, transition updates become infrequent and the transition matrices form a Cauchy sequence. Neither implication follows. For example, s(t) = 1/(t+1) reaches a new minimum at every step, although its successive decreases tend to zero. Moreover, Eq. (5) uses a fixed step size, and small changes in entropy do not bound changes in the transition targets obtained from hard predictions. A moving average of alternating target matrices can continue to oscillate. Lemma 1 is valid, but it establishes a property of the running minimum of any bounded sequence; it does not establish convergence of the classifier or the transition matrix. The central convergence claim therefore needs additional assumptions and a valid argument.

  2. The binary noise experiments need clarification. Figure 1 defines uniform noise as flipping a label uniformly to another class. For Breast and Pneumonia, which have two classes, this gives the same corruption distribution as cyclic flipping and the custom mapping in Table 3. Yet Table 1 reports a large difference between these settings. On Pneumonia at 50% noise, SLT obtains 83.90 F1 under UN but 47.44 under CF/CM; the other methods also show substantial differences. Please provide the actual corruption matrices and the realized fraction of changed labels. In particular, does the UN implementation sample from all classes, including the original class? If so, the effective corruption rate differs from the definition in Figure 1 and from the nominally matched CF/CM settings.

  3. It is unclear why the entropy trigger should improve the accuracy of the noise-transition estimate. Equation (4) estimates P(noisy label = k | predicted label = j), whereas forward correction ordinarily uses P(noisy label = k | clean label = j). These quantities differ when the classifier makes mistakes. Lower entropy alone does not resolve this: a model that confidently fits the corrupted training labels would produce an identity matrix through Eq. (4), even when the true corruption is nonzero. Likewise, a more diagonal estimated matrix in Figure 7 does not by itself demonstrate more accurate noise estimation. Since the synthetic corruption matrices are known, reporting transition-estimation error over training would directly test this part of the motivation.

Questions for authors

See above.

Reviewer_FQuJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a Quantum Neural Network for learning from noisy labels. Arguing that intrinsic natural smoothness is one of the properties of QNNs, the paper highlights that learning from noisy labels is even more challenging than with classical CNNs. To this end, the paper proposes supermartingale-based Label Transition (SLT) and utilizes its monotonic behavior to identify a stable transition-matrix refinement step within an anchor-free loss correction framework. The framework is evaluated on five different MedMNIST datasets: BreastMNIST, PneumoniaMNIST, RetinaMNIST, DermaMNIST, and BloodMNIST. The experimental results show that the proposed method outperforms the existing methods used for comparison.

Critical strengths
  1. QNN is a promising and growing research area.
  2. The paper is well motivated and easy to follow.
  3. The method is explained with pseudocode, and the ideas are also supported by theory. .
Critical weaknesses

QNN is not my area of expertise; therefore, I might not be able to rigorously assess the method. After carefully going through the paper, the method seems reasonable to me. However, I find the evaluation limited to small-scale datasets, and it is a bit difficult for me to gain confidence in the generalization of the proposed pipeline to large-scale and naturally noisy datasets. Therefore, comparing performance on large-scale, naturally noisy datasets in the natural domain (for example, Clothing1M) would give a better idea of the method’s generalization.

Questions for authors

Is there any specific reasons why you limited your experiments only on medical datasets? The methods seems to be generic, and evaluating it on the diverse dataset would give more confidence on the method. Would you be able to evaluate on existing datasets such as Clothing1M and compare with the baselines?

Reviewer_mdkYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper trains quantum neural networks for medical image classification, where data are scarce and the labels are noisy. The problem it targets is estimating the label-noise transition matrix without reliable anchor samples. SLT applies forward loss correction and refreshes the transition matrix whenever the mean predictive entropy hits a new low. Each refresh blends a fresh estimate (predicted classes against observed labels) into the previous matrix by moving average. Experiments run on medical image datasets with both synthetic and real-world label noise.

Critical strengths
  • Scarce data and annotation errors are a genuine obstacle in medical imaging, so adapting noise correction to the quantum setting is a worthwhile problem to work on.
  • The update rule is simple and drops into a hybrid quantum-classical classifier without much machinery: an entropy-based trigger, then moving-average smoothing of the transition matrix.
  • Experiments span several datasets and noise settings, with repeated runs and reported variability. The extra analyses of backbone choice, update strategy, temperature scaling, and runtime help put the reported gains in context.
Critical weaknesses
  • The quantum advantage is overclaimed. A trainable classical classifier sits after the measurement, so the Born rule alone says nothing about whether the final class probabilities are smooth. To credit the circuit rather than raw capacity or a better-tuned optimizer, I would want parameter-matched classical models, comparable tuning budgets, and prediction-entropy and calibration curves. Simulation results also tell us nothing about hardware.
  • The CheXpert adaptation is underspecified. The formulation is single-label, but CheXpert is multi-label. Please state how transition matrices are defined per condition, how missing labels are handled, how decision thresholds are chosen, and how patient-level splits are built.
  • No experiment shows that entropy-triggered updates actually improve the estimate or keep it stable over time. Entropy curves and a handful of matrix snapshots are indirect at best. The synthetic noise transitions are known, so please report estimation error against them and track how the matrix moves across training.
Questions for authors

please refer to weakness 2-3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.