Abstract
Detecting ultra-rare objects (frequency -6) without positive annotations remains a critical challenge in computer vision. We propose a Synthesis-Prior Driven paradigm that eliminates the need for real positive samples. Our core insight is that models trained on statistically and perceptually indistinguishable synthetic data can generalize effectively to real-world environments. To achieve this, we introduce a Differentiable Morphological Modeling framework, which embeds explicit physical and geometric priors into a differentiable synthesis engine. Unlike deep learning generative models, this approach ensures high-fidelity imagery with perfectly accurate, objective annotations. We validated this paradigm on the safety-critical task of detecting Circulating Genetically Abnormal Cells (CACs). Perceptual assessments confirmed that synthetic data is indistinguishable from clinical imagery (expert accuracy ≈ 50%). Consequently, our model, trained exclusively on synthetic data, achieved state-of-the-art performance on real patient samples (98.43% accuracy, 97.40% sensitivity), outperforming both human experts and existing automated methods. Furthermore, the model demonstrated exceptional robustness across a multi-center cohort (CV < 0.8%) and achieved high-precision zero-shot transfer in industrial defect inspection (DAGM 2007). This study establishes a principled methodology for ultra-rare object detection, offering broad applicability across domains with definable geometric priors.