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

Jun Zhuang, Mohammad Al Hasan, Yiyu Shi, Chaowen GuanarXiv 2026 · arXiv

TL;DR

This paper introduces Supermartingale-based Label Transition (SLT), a novel anchor-free loss correction framework designed to make Quantum Neural Networks (QNNs) robust against noisy labels in small-scale medical image datasets. By modeling the network's entropy reduction process as a mathematical supermartingale, SLT leverages the unique "natural smoothness" of QNNs to stably update noise transition matrices without relying on hard-to-find clean anchor samples.

What Problem Does This Paper Address?

Medical image classification faces a dual challenge: datasets are often small due to the high cost of data acquisition, and labels are frequently noisy due to inter-observer variability among clinical experts. While Quantum Neural Networks (QNNs) show great promise in these limited-data regimes due to their high parameter efficiency, their behavior under label corruption has been poorly understood.

Traditional methods fix noisy labels by estimating a "noise transition matrix" using highly confident clean samples called anchor points. However, QNNs possess an intrinsic "natural smoothness" rooted in the Born rule of quantum mechanics. Instead of producing the sharply peaked, overconfident distributions typical of classical deep neural networks (DNNs), QNNs output smoother probability distributions. This smoothness acts as an excellent implicit regularizer but presents a catch-22: it obscures the explicit anchor points needed to estimate how labels are being corrupted.

Methods

The authors propose SLT (Supermartingale-based Label Transition), an anchor-free framework that turns QNN smoothness from a limitation into a stabilizing asset.

  1. Supermartingale Modeling: Instead of updating the noise transition matrix erratically whenever the model seems locally confident, SLT tracks the normalized predictive entropy across the dataset. It models the historical minimum of this entropy as a supermartingale—a stochastic process that, on average, decreases or stays constant over time.
  2. Entropy-Triggered Updates: The noise transition matrix is only updated when the current predictive entropy drops below its historical minimum. This ensures that updates are triggered dynamically only when the model achieves a new, stable threshold of global confidence.
  3. Moving Average Refinement: To mitigate noise-driven oscillations during training, the global transition matrix is updated smoothly using a moving average rule based on the current low-entropy predictions.

The overall workflow of the training and refinement process is illustrated below:

Figure 2: Overall workflow of our proposed frame- work, SLT. Solid arrows denote the main training process, and dashed arrows indicate the entropy- driven transition-matrix refinement.

Datasets

The framework was rigorously evaluated on two main benchmarks:

  • Five MedMNIST datasets: BreastMNIST, PneumoniaMNIST, RetinaMNIST, DermaMNIST, and BloodMNIST. These public datasets represent small-scale medical imaging tasks across various modalities (ultrasound, X-ray, fundus camera, dermatoscope, and microscope) injected with synthetic noise (Uniform, Cyclic-flipping, and clinically-inspired Custom-mapping noise).
  • CheXpert: A large-scale real-world chest radiograph dataset containing naturally occurring uncertain/noisy labels. The authors downsampled it to 10% to simulate a data-scarce environment.

Key Insights and Ideas

  • Embracing Quantum Smoothness: Rather than forcing QNNs to behave like classical networks (e.g., by using temperature scaling to artificially sharpen outputs), SLT exploits their gradual confidence growth. This steady entropy reduction prevents the model from prematurely overfitting to corrupted labels.
  • Provable Convergence: By formalizing the optimization via supermartingales, the authors provide a mathematical proof (utilizing Doob's Forward Convergence Theorem) showing that the transition-matrix updates form a Cauchy sequence and are guaranteed to converge to a steady state.
  • Extreme Parameter Efficiency: The QNN backbone achieved state-of-the-art robustness using a fraction of the parameters required by classical architectures. For instance, the QNN used just 586 trainable parameters, compared to 2,146 for a shallow classical network and 23,266 for a ResNet variant.

Key Results

  • Robustness Under Severe Noise: On BreastMNIST and PneumoniaMNIST with 50% uniform noise, SLT achieved macro F1 scores of 52.24% and 83.90% respectively, significantly outperforming standard cross-entropy and classical anchor-free baselines like VolMinNet and CCR.
  • Real-World Superiority: On the real-world CheXpert dataset, SLT achieved a state-of-the-art Area Under the Curve (AUC) of 89.85% and an F1 score of 68.94%, outperforming domain-specific medical imaging methods such as VisualCheXbert and BoMD.
  • Backbone Comparisons: As shown in the ablation studies, classical networks (SNN and DNN) suffered from early-stage overconfidence, causing their predictive entropy to collapse prematurely and degrading performance when paired with SLT, whereas QNNs naturally thrived.
Ablation Study

Limitations and Future Work

The primary limitation acknowledged by the authors is the high computational cost of simulating quantum circuits on classical hardware, which remains a standard bottleneck in Quantum Machine Learning (QML).

For future work, the authors plan to evaluate the SLT framework under more complex, instance-dependent real-world noise patterns and expand its validation across a broader array of medical imaging modalities.

Summary updated

Abstract

Noisy-label learning in small-scale medical image classification is challenging and hinders the superiority of deep neural networks. Recent studies suggest that quantum neural networks (QNNs) have shown potential in limited-data regimes, yet their use for noisy-label learning remains under-explored. 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 consistently improves QNN-based classification and stably outperforms classic noise-label learning baselines under synthetic and real-world label noise.
BibTeX
@misc{zhuang2026slt,
  author = {Jun Zhuang and Mohammad Al Hasan and Yiyu Shi and Chaowen Guan},
  title = {{SLT: Robust Quantum Neural Networks for Noisy-Label Medical Image Classification via Supermartingale-based Label Transition}},
  year = {2026},
  eprint = {2607.16293},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.16293},
}
Figure