In medical diagnostics, integrating quantum computing with deep learning offers potential improvements in sensitivity and computational efficiency. This study introduces two models: LiteSpineNet, a streamlined CNN, and HQ-LiteSpineNet, its hybrid quantum counterpart that incorporates a novel seven-qubit variational feature extraction block for multiclass spine lesion classification. Using the VinDr-SpineXR dataset of 12,000 X-ray images across eight classes, we show that both models achieve strong performance while requiring substantially fewer parameters than state-of-the-art methods. LiteSpineNet attains 76.81% test accuracy, 79.69% precision, 76.81% recall, 77.92% F1, and 89.75% AUROC. HQ-LiteSpineNet achieves 72.43% accuracy, 80.40% precision, 72.43% recall, 75.46% F1, and 89.95% AUROC. Importantly, HQ-LiteSpineNet improves minority class detection, raising recall by +6.7% and F1 by +4.4% compared to the classical baseline. Statistical significance was confirmed through paired t-tests (p<0.0001), supported by 99% confidence intervals, external validation, and decision curve analysis, which showed net benefit gains in minority classes. Computational cost analysis highlights the exponential overhead of quantum simulators (O(2ⁿ)) but emphasizes the linear scalability of real quantum hardware (O(n)). These findings establish LiteSpineNet as an efficient baseline and HQ-LiteSpineNet as a promising quantum-enhanced alternative, advancing AI-driven diagnostics by combining efficiency with improved sensitivity in underrepresented lesion categories.