acceptodds
Withdrawn from ICLR 2027

Quantum-Assisted Prototype-Guided Refinement for Robust Medical Image Classification under Label Noise

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Labels of medical images are often corrupted by annotation errors, diagnostic ambiguity, and uncertainty in report-derived findings, and deep networks trained on such labels memorize the errors. Prototype-based label correction can repair corrupted supervision, but only in an embedding where classes are separated and concentrated; otherwise erroneous assignments contaminate the prototypes and reinforce the errors they are meant to correct. We propose HQProNet, which couples trusted-anchor prototype refinement with a compact variational quantum circuit (VQC) that defines the matching geometry. Prototypes are initialized from a small audited anchor set, temperature-scaled cosine similarities yield refined soft targets, and confidence-gated moving-average updates adapt the prototypes during training. The VQC maps a unit-norm 256-dimensional amplitude vector to eight Pauli- expectations, each a trainable quadratic form with eigenvalues in , using 120 rotation angles. A conditional analysis shows that prototype matching recovers the clean label when the number of trusted anchors, the within-class covariance, and the cosine margin jointly satisfy an explicit condition. On six MedMNISTv2 datasets under symmetric and pairwise corruption, HQProNet attains the best macro-F1 in 34 of 36 conditions, and on CheXpert it is best on all six metrics under two uncertainty policies. With the prototype mechanism fixed, replacing the VQC by linear, nonlinear, or parameter-matched classical projectors lowers macro-F1 by at least 2.2 points in every setting and reduces the cosine margin and refinement accuracy: prototype refinement is the primary source of robustness, and the VQC supplies the most favorable geometry among the projectors tested.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Learning with noisy labels, medical image classification, variational quantum circuits, hybrid quantum–classical learning, representation geometry

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_zzFeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

HQProNet addresses label noise in medical image classification. A CNN feature (80-d) is projected to a unit 256-d vector, amplitude-encoded into an 8-qubit, 5-layer VQC, and the 8 Pauli-Z readouts define the space for trusted-anchor prototype refinement. Prototypes are initialized from a small audited clean set (5% of validation), temperature-scaled cosine similarities give soft targets mixed with observed labels, and prototypes are updated by confidence-gated EMA. Theory: Bernstein concentration of anchor means plus a cosine-margin condition for correct matching, a mixed-target noise identity, and an EMA drift bound. Validation: six MedMNISTv2 datasets under 10–50% symmetric/pairwise synthetic noise, a 10% CheXpert subset under two uncertainty policies, and a controlled study replacing the VQC with linear/MLP/budget-MLP projectors. HQProNet is reported best in 34/36 MedMNIST conditions and on all CheXpert metrics, with ~2–3 point gains over the best classical projector.

Critical strengths
  1. Careful, honest scoping of the quantum contribution. The authors explicitly disclaim quantum advantage and use a controlled ablation (Tab. 3, 5) to show that prototype refinement, not the VQC, is the main source of robustness (40.2→75.5 with a linear projector). This makes the paper's claims checkable and is the main reason it is not a clear reject on framing alone.
  2. Label-repair diagnostics. Reporting refinement accuracy, repair precision/recall and retention alongside macro-F1 (Tab. 16–17) shows the method corrects labels rather than relabeling indiscriminately. This is a useful evaluation practice.
Critical weaknesses
  1. The main comparison is unfair, and the VQC's contribution is not isolated.

    • HQProNet uses a clean, audited anchor set, while all noisy-label baselines (DivideMix, Co-teaching, NQNN, DUE-Net) use none. The paper's own ablation shows a linear projector with the same anchors already gets most of the gain. The headline "best in 34/36" therefore largely reflects access to trusted data.
    • Trusted-data baselines are absent: GLC, MW-Net/L2RW, DMLP, and simple fine-tuning on the anchors.
    • The VQC readout (Eq. 4) is a classical, constrained quadratic form z⊤Mjzz^\top M_j z, but no classical quadratic-form projector is tested.
    • Hyperparameters were tuned on HQProNet and reused for the controls.
    • The resulting ~2-point gap over 3 seeds, without significance testing, does not support the claim that the VQC supplies a better geometry.
  2. Several reported numbers show patterns that undermine trust in the results.

    • Pairwise-noise entries in Tab. 1 are ≈0.958× the symmetric entries for every method.
    • CheXpert quantum-baseline values in Tab. 2 coincide with MedMNIST entries in Tab. 1 (e.g., QTNet 13.1/12.0, QCCNet 13.6/13.2, A-HQCNN 22.6/25.4).
    • Several AUCs are far below 0.5 (12.0, 16.1).
    • Gaps between projectors are nearly constant across all datasets (Budget-MLP is 0.6–0.7 below MLP in nearly every cell; the margin gap is exactly 0.032 under both noise types).
    • Until these are explained with per-seed logs, the empirical evidence cannot be relied upon.
Questions for authors

Clarifications (could change my assessment):

  1. Why are the pairwise results in Tab. 1 ≈0.958× the symmetric results for every method? Why do the CheXpert baseline rows in Tab. 2 match MedMNIST entries in Tab. 1? How were AUCs well below 0.5 obtained? Please provide per-seed logs.
  2. Were the pairwise-noise and CheXpert baseline experiments run independently? Were baseline hyperparameters tuned per method?
  3. How does HQProNet compare with GLC, MW-Net, DMLP, and fine-tuning on the same 5% anchors?
  4. With identical hyperparameter tuning, what does a classical trainable quadratic-form projector (e.g., low-rank symmetric MjM_j with bounded spectrum) achieve?

Suggestions (would not by themselves change my rating): 5. Report a clean-label upper bound and evaluate whether the Eq. 13 recovery condition actually holds for the realized anchor counts.

Reviewer_EJ47OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose HQProNet to help address noisy labels in multi-class and multi-label tasks, specifically in medical imaging which is especially prone to annotation errors, diagnostic ambiguity, and report uncertainty. They use an 8-qubit, 5-layer variational quantum circuit to obtain features which are then used for cosine matching against class prototypes which are initialized from a small subset of trusted samples. In addition to the observed labels, prototype-derived soft targets and confidence-gated updates to adapt the prototypes are used during training. The authors report that prototype refinement helps the most but the variational quantum circuit features improve feature geometry and label-refinement accuracy. They validate HQProNet using 6 of the MedMNISTv2 datasets with symmetric and pairwise label corruption and get the best macro-F1 in 34 of 36 conditions, and on CheXpert, they get the best on all 6 of 6 conditions under two report-uncertainty policies.

Critical strengths
  1. The authors provide a good set of experiments across six MedMNISTv2 datasets with symmetric and pairwise noise at different levels and CheXpert under two report-uncertainty pliticies and achieve the best results in nearly all cases against strong and recent baselines.
  2. The authors also provide a strong set of ablations of components, circuits, projectors, and trusted-anchors.
Critical weaknesses
  1. It would have been great to see any results on medical imaging datasets with real noise such as MIMIC-CXR for example. Great to see results on simulated noise, but real-world instance noise, which is much more common in medical imaging, may be a good indicator of the strength of HQProNet.
  2. The proposed method needs a bit of auditing to find some correctly labeled samples. Although the authors show that more anchors help, it is unclear whether/how much underrepresentative anchors, rare cases, etc. affect the method.
Questions for authors
  1. Is the “quantum” part of the proposed method significantly better than a very similar “non-quantum” alternative?
  2. Any reason only half of the MedMNISTv2 datasets were used?
Reviewer_WzWSOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes HQProNet, a framework for medical image classification with noisy labels that combines trusted-anchor prototype refinement with a variational quantum circuit (VQC). A classical encoder extracts image features, and an eight-qubit VQC maps them into an embedding where similarities to class prototypes generate refined soft labels. These labels are combined with the original supervision during training, while confident assignments update the prototypes through an exponential moving average. The authors provide a conditional analysis linking refinement reliability to anchor count, class concentration, and separation, and report improvements across six MedMNIST datasets and CheXpert. Their central claim is that prototype refinement provides most of the robustness, while the VQC improves the feature geometry used for correction; all quantum experiments use classical simulation.

Critical strengths
  1. Relevant problem and clear motivation. The paper addresses noisy supervision in medical image classification and identifies the dependence of prototype-based label refinement on the separation and concentration of the feature representation.
  2. Controlled component comparisons. The inclusion of linear, nonlinear, and approximately parameter-matched classical projectors within the same trusted-anchor refinement pipeline helps distinguish the contribution of prototype refinement from that of the VQC.
Critical weaknesses
  1. The prototype-update mechanism cannot operate as described. With cosine softmax at T=1T=1, PathMNIST’s maximum possible confidence is approximately 0.48, below the threshold 0.7. Thus, prototypes cannot update on five of six datasets, contradicting the interpretation of the EMA sensitivity experiments.
  2. The benefit of the quantum component is insufficiently established. The VQC computes structured quadratic features, but the classical controls include only linear and MLP projectors. Classical quadratic or bilinear alternatives are needed to identify what the circuit contributes.
  3. The theory does not establish reliability of the actual training procedure. It analyzes a fixed representation and clean anchor means, whereas training uses an evolving representation and prototype bank. It also provides no VQC-specific explanation for improved geometry.
Questions for authors

Confidence gating: With cosine scores in ([-1,1]) and (T=1), the maximum softmax confidence on nine-class PathMNIST is approximately 0.48, below (\tau=0.7). How can samples enter the prototype-update set? Please clarify whether the implementation uses a different confidence definition or additional logit scaling, and report admission rates across epochs.

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

This paper proposes HQProNet, which combines a variational quantum circuit (VQC) with trusted-anchor prototype refinement to improve medical image classification under label noise. The VQC projects image features into a low-dimensional representation space, where class prototypes are used to generate refined soft labels through cosine similarity. The prototypes are subsequently updated using high-confidence predictions. Experiments on six MedMNISTv2 datasets and CheXpert show improvements over several existing noisy-label learning methods.

Critical strengths

(1) The experimental evaluation is fairly comprehensive. The authors consider different noise types and corruption rates, and provide ablations that separate the contributions of prototype refinement and the quantum projection module. The analysis of class margins, within-class covariance, and label correction accuracy also helps explain the observed results.

(2) The problem itself is relevant to medical imaging, where annotation errors and uncertain diagnoses are common. Using a small set of trusted samples to correct noisy supervision is a practically meaningful direction.

Critical weaknesses

(1) The necessity of the quantum component remains questionable. According to Table 3, under 50% symmetric label noise, adding prototype refinement with a classical linear projector already improves the average Macro-F1 from 40.2% to 75.5%, compared with 78.9% for the complete model. Although several classical projectors are evaluated, the paper does not include stronger alternatives that more closely match the mathematical structure of the VQC, such as quadratic feature mappings or bilinear networks. The results therefore establish the effectiveness of prototype refinement more convincingly than the unique contribution of quantum computation.

(2) The comparison with existing noisy-label methods is not entirely fair. HQProNet uses an additional set of trusted labels, selected from 5% of the validation samples per class, to initialize its prototypes. Several competing methods do not receive equivalent clean supervision. This difference could account for part of the reported improvement.

(3) The evaluation remains somewhat idealized. Most experiments rely on synthetic label corruption, while the CheXpert experiments use only a 10% subset. There is no independent cross-institutional validation. Moreover, all quantum circuits are simulated on classical hardware, so the work does not establish any computational or practical advantage from quantum execution.

Questions for authors

Two additional experiments would be particularly important.

First, does HQProNet still outperform existing noisy-label learning methods when all methods have access to the same trusted-label budget?

Second, how does the VQC compare with classical quadratic or bilinear feature mappings under matched parameter counts and computational budgets? Such comparisons would help establish whether the observed gains arise from the proposed quantum representation or from a more general nonlinear transformation.

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

Related papers

Discussion (0)

Sign in to comment.