Are Hybrid Quantum Medical Image Classifiers Robust to NISQ Noise? A Leakage-Controlled Calibration and Finite-Shot Audit

Le résumé fourni par la source

This repository contains the supplementary material, detailed experimental results, statistical analyses, and reproducibility files supporting the study “Are Hybrid Quantum Medical Image Classifiers Robust to NISQ Noise? A Leakage-Controlled Calibration and Finite-Shot Audit.” The study investigates the robustness of hybrid quantum–classical medical image classifiers under simulated noisy intermediate-scale quantum (NISQ) conditions. Experiments were conducted using BreastMNIST and PneumoniaMNIST, with frozen ResNet18 feature extraction followed by train-only preprocessing and PCA-based dimensionality reduction. Two quantum learning approaches—a variational quantum classifier (VQC) and a ZZ-feature-map quantum kernel classifier (QSVM-ZZ)—were evaluated alongside classical machine-learning baselines. The supplementary package provides detailed results for depolarizing, amplitude-damping, phase-damping, bit-flip, and combined quantum noise channels across multiple noise levels. It also includes prediction-drift analyses, calibration and discrimination metrics, runtime information, and finite-shot sensitivity experiments covering 64, 128, 256, 512, and 1024 shots. The finite-shot analysis contains 108,000 evaluation records generated across repeated Monte Carlo simulations. The repository includes: Publication-ready supplementary material containing Tables S1–S7 Leakage-audit and cleaned-split information Classical baseline results Complete VQC and QSVM-ZZ channel-noise results Holm-corrected statistical-test outputs Prediction label-flip and probability-drift analyses Finite-shot sensitivity summaries and complete finite-shot results Runtime comparisons Experimental configuration and run-state files The files are provided to support transparency, reproducibility, independent verification, and secondary analysis of the reported experiments. The original medical imaging datasets are not redistributed in this repository; BreastMNIST and PneumoniaMNIST are publicly available through MedMNIST.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.