Squid–transmon quantum hardware simulation with deep learning for pancreatic radiotherapy image classification

We present a hybrid quantum–classical framework that couples a parameterized, hardware-imperfection-aware SQUID–Transmon surrogate simulation with deep learning to classify pancreatic radiotherapy CT/CBCT images from the collection. The classical backbone uses a pretrained with bidirectional Long Short-Term Memory layers; the quantum component is a 4-qubit, shallow-depth circuit whose input angles are modulated by a dynamically computed error-mitigation factor derived from the gradient of an effective potential. We perform a grid search over three hardware-relevant parameters: a rescaling factor R R (an amplitude/drive scaler in our surrogate) that modulates the effective potential and controls the input rotation amplitude, a secondary coupling strength λ λ , and the bare Josephson energy EJ0 E J 0 . In a 10-class subset experiment (top-dose patients), the hybrid model exhibits a performance maximum when $$R=0.95$$ R = 0.95 , reaching a test accuracy of up to 0.95 across several (λ , EJ0) ( λ , E J 0 ) settings. In this work, deviations of R R , λ λ , and EJ0 E J 0 are treated as quasi-static hardware imperfections (systematic calibration offsets and slow drift) rather than as stochastic noise channels. We interpret this optimum as a balance between expressivity and over-rotation/leakage in our simulator when the effective drive is slightly reduced. To benchmark consistency on the full dataset, we additionally report a window-level evaluation over all 40 patients with the classical backbone alone: a single 80/10/10 split yields train=0.959 train = 0.959 , val=0.960 val = 0.960 , test=0.909 test = 0.909 accuracy (windows); stratified threefold cross-validation achieves 0.876 ± 0.005 0.876 ± 0.005 mean validation accuracy. Overall, our results indicate that such hardware-imperfection-aware hybrid models can be competitive with strong classical baselines while offering a physics-grounded knob for hardware-aware calibration; we discuss modeling assumptions and limitations (e.g., simplified hardware-imperfection model, connectivity, and task definition) to avoid overclaiming clinical readiness.