Pancreatic ductal adenocarcinoma (PDAC) exhibits profound chemoresistance largely driven by its dense stromal architecture, necessitating the use of 3D organoid models for accurate preclinical drug screening. However, conventional viability assays (e.g., ATP-based luminescence or immunofluorescence) are destructive, label-dependent, and fundamentally limited in capturing the spatial heterogeneity of drug penetration. Here, we present a novel label-free high-content screening (HCS) platform leveraging an unsupervised anomaly detection model (SK-RD4AD) to evaluate the spatial efficacy of anti-cancer drugs using simple phase-contrast microscopy. The model autonomously segments the organoid region of interest (ROI) from background noise and maps morphological degradation patterns as stress gradients. By translating these AI-derived anomaly scores into a clinically relevant total drug response spectrum (TDRS), the platform precisely quantified drug efficacy without chemical labeling. Our results demonstrated a highly accurate stratification of candidate samples into high response (TDRS >70%, showing widespread structural breakdown) and low response (TDRS <10%, intact spheroids indicating strong resistance) categories. This convergence of computer vision and pharmacology overcomes the critical bottlenecks of 3D organoid screening, offering a highly reproducible, cost-effective, and non-invasive pipeline for precision medicine and next-generation drug discovery.