A Geometry-Guided Generative Framework for High-Fidelity Dermatological Analysis under Data Scarcity

Current methodologies in fine-grained visual analysis—specifically in biomedical domains—face a critical dichotomy: deep learning models require massive datasets that are often unavailable for rare pathologies, while traditional geometric reconstruction lacks semantic interpretive capabilities. This research proposes Geo-Synth, a hybrid framework that utilizes dense depth priors to guide synthetic data generation, creating a robust pipeline for analyzing complex surface anomalies such as dermatological conditions with extremely limited ground-truth data. We introduce a dual-stream architecture fusing Total Variation optimized depth estimation with Generative Adversarial Networks (GANs). Our approach employs an adaptive random-walk algorithm to refine surface normals and a physics-informed generator that conditions synthesis on geometric topology. Extensive experiments on benchmark dermatological datasets demonstrate that Geo-Synth reduces the need for labeled data by 40% while improving segmentation Intersection-over-Union (IoU) scores by 12.4% compared to state-of-the-art purely pixel-based augmentation methods. Furthermore, theoretical analysis confirms that our geometric coupling ensures Lipschitz continuity across the generated manifold, preventing mode collapse common in low-data regimes.