HypDeformNet: Edge-Deployable Deep Architecture with Jacobian-Stable Hyperbolic Deformation and Lipschitz Distillation for Immunotherapy Response Prediction

Predicting response to PD-L1 immunotherapy in metastatic cancer from computed tomography (CT) is challenging in data-scarce, imbalanced cohorts requiring generalization from few annotated lesions and subtle pre-treatment phenotypes. We present HypDeformNet, an edge-deployable framework combining morphology-conditioned hyperbolic deformations with Jacobian stability constraints and Lipschitz-controlled online teacher-student distillation in hyperbolic space. Experiments on 44 metastatic urothelial carcinoma (mUC) patients with four outcome classes (Complete Response (CR), Partial Response (PR), Stable Disease (SD), Progressive Disease (PD)) under patient-stratified validation achieve patient-level accuracy of 90.9% (teacher) and 88.6% (student), with the distilled model retaining 97.5% teacher performance using 3.8× fewer parameters. Lesion-level upper bounds reach 94.7% and 93.8% respectively. Ablations confirm geometry-aware, morphology-preserving augmentation as central to robust prediction. The compact student model enables real-time on-scanner deployment for embedded CT immunotherapy response assessment.