کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

DepthFusionSAM: Training-Free Colonoscopy Polyp Segmentation via Retrieval-Guided Prompting

نویسندگان :

Ali Tamizifar¹ Shakiba Berenjkoub² Nader Karimi³ Shadrokh Samavi⁴

1- Department of Electrical and Computer Engineering, Isfahan University of Technology 2- Department of Electrical and Computer Engineering, Isfahan University of Technology 3- Department of Electrical and Computer Engineering, Isfahan University of Technology 4- Department of Computer Science, Seattle University

کلمات کلیدی :

Polyp segmentation،Colonoscopy،SAM3،Training-free learning،SegGPT،DepthFusionSAM

چکیده :

Accurate colorectal polyp segmentation is vital for early cancer detection, yet existing deep learning methods often require extensive training or fine-tuning to handle the high variability and complex backgrounds of colonoscopy images. This paper introduces DepthFusionSAM, a novel three-stage, training-free framework for automated polyp segmentation. The framework first utilizes DINOv2 embeddings and FAISS models to retrieve visually similar template images that provide query-specific priors. In the second stage, SegGPT generates multiple candidate masks, which are then re-ranked using semantic consistency and scale-based priors. Finally, a median box prompt is fed into SAM3, augmented by DepthFusion—a depth-guided feature gating mechanism derived from a pre-trained depth estimation model—to enhance robustness against specular highlights and low-contrast boundaries. Evaluated on benchmark datasets including Kvasir-SEG and PolypGen, DepthFusionSAM achieves competitive performance without any dataset-specific optimization, offering a scalable and efficient solution for real-time clinical workflows.