| Abstract |
Lung cancer is the leading cause of cancer-related death worldwide, and automated malignancy classification of pulmonary nodules from CT scans is a high-value clinical problem. This work investigates whether a Quantum Extreme Learning Machine can extract more discriminative representations of radiomic data than classical methods while using a fraction of the input features. A 10-qubit circuit with redundant angle encoding and a random transverse-field Ising reservoir was evaluated on 526 nodules from the LIDC-IDRI dataset, with 56 features reduced to 4 via ANOVA selection. The QELM achieved the highest AUC (0.971) and accuracy (91.5%) across all models, outperforming five classical baselines, including Logistic Regression, MLP, and Random Forest, that were trained on all 56 features. The results demonstrate that the 512-dimensional Hilbert space of the Ising reservoir generates nonlinear feature representations sufficient to surpass classical performance with a 14x reduction in input dimensionality, establishing feature efficiency as the primary empirical finding of this study. |
