Evaluation integrity in hybrid quantum-classical skin lesion classification: A data-leakage audit and lesion-level correction on HAM10000

This paper presents a systematic ablation and evaluation integrity study of hybrid quantum-classical neural networks for seven-class skin lesion classification on the HAM10000 dermoscopy dataset. We study the effect of the evaluation protocol on a hybrid quantum-classical pipeline by comparing two partitioning schemes on the same model and data. The first is a leakage-prone scheme that partitions the pool of augmented images, and the second is a verified scheme that partitions original lesion identities. Under the leakage-prone scheme, augmented variants of the same physical lesion appear on both sides of the split. An audit found that 1,095 of 4,810 lesions (22.8%) crossed the split boundary, artificially inflating all previously reported metrics. We rebuild the pipeline with a verified lesion-level partition consisting of 4,152 training, 522 validation, and 517 original test images. The training set is augmented to 10,503 images while ensuring zero image-level and lesion-level overlap. We then re-run every experiment across five random seeds with per-class Wilson confidence intervals and paired significance tests on a fixed held-out test set that is never used for tuning. Under the corrected protocol, the strongest configuration features a ResNet50 backbone with concatenation fusion, focal loss, and test-time augmentation. This setup reaches 78.41% ± 0.74 test accuracy and a macro F1 of 0.804 against a verified baseline of 76.17% ± 1.91 . However, under the leakage-prone protocol, this exact same pipeline yields 94.57%. A parameter-matched classical control, in which a 36-parameter multilayer perceptron replaces the variational quantum circuit (VQC), is statistically indistinguishable from the hybrid model (78.14% versus 78.41%, p = 0.81). Additionally, a controlled fusion comparison shows that concatenation is not significantly better than weighted fusion. We therefore report no detectable advantage from the quantum branch at this scale as a controlled negative result rather than a failure, and we characterize VQC trainability under noiseless and depolarizing simulation. Finally, a split-audit survey of recent quantum-hybrid HAM10000 studies indicates that reported accuracies above 90% are associated with undisclosed or leakage-prone evaluation protocols. The contribution of this work is methodological: a reproducible leakage audit, a verified correction protocol, an honest multi-seed baseline, a parameter-matched classical control, and evidence that evaluation protocols rather than architectures drive the headline differences reported in this literature.

• Data leakage
• dermoscopy
• evaluation protocols
• feature fusion
• HAM10000
• hybrid quantum-classical neural networks
• quantum machine learning
• reproducibility
• skin lesion classification
• variational quantum circuits

Jamaludin, M. R. A., & Kim, C. S. (2026). Evaluation integrity in hybrid quantum-classical skin lesion classification: A data-leakage audit and lesion-level correction on HAM10000. IEEE Access, 14, 101415–101428.

@article{jamaludin2026ham10000,
  author = {Jamaludin, M. Rifqi Azkiya and Kim, Chang-Soo},
  title = {Evaluation integrity in hybrid quantum-classical skin lesion classification: A data-leakage audit and lesion-level correction on HAM10000},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {101415--101428},
  doi = {10.1109/ACCESS.2026.3707673}
}

Copy BibTeX