Near-term quantum machine-learning studies often report performance values without a fully specified split, feature-selection protocol, or comparison with equally constrained classical baselines. This paper presents a reproducible simulation-based hybrid quantum convolutional neural network (QCNN) for binary breast-cancer diagnosis using the Wisconsin Diagnostic Breast Cancer dataset. Binary particle swarm optimization (PSO) reduces 30 fine-needle-aspirate features to 14 candidates, which are ranked and mapped to 2-, 4-, 6-, and 8-qubit circuits. An RY angle encoding and a ZZ entangling map are evaluated under the same parameter-sharing QCNN, followed by a class-weighted logistic readout. Across three circuit initializations, the 8-qubit RY model achieved 92.98% accuracy, 93.25% precision, 87.30% recall, a 90.16% F1-score, 96.30% specificity, and an ROC AUC of 0.9850. The 4-qubit model retained 91.52% accuracy and 0.9805 AUC. The ZZ encoding was less stable. Classical RBF-SVM and logistic-regression baselines using the same eight features reached 95.61% accuracy; therefore, no quantum advantage is claimed. The study provides an honest resource-performance benchmark and identifies the need for noise-aware hardware validation.
 

breast cancer, particle swarm optimization, quantum convolutional neural network, quantum feature encoding, hybrid learning.