From Single-Round to Sequential: Building Stateful Interactive Segmentation with SegVol and GRU Corrector

Chuanyi Huang, Jinlong Huang, Lisheng Wang

Keywords: Interactive Medical Image Segmentation; Sequential State Modeling; Uncertainty-Driven Refinement; GRU-Based Correction

TL;DR: We introduce a stateful interactive segmentation framework using SegVol and GRU Corrector, enabling efficient and precise multi-round refinement for medical imaging with fewer user interactions.

Abstract: Medical image segmentation has advanced significantly with foundational models like Segment Anything Model (SAM), but real-world clinical applications face challenges due to heterogeneous imaging protocols, small irregular structures, and inefficient interactive refinement. Existing methods lack memory-aware processing, struggle with modal constraints, and exhibit poor generalization. We propose "From Single-Round to Sequential: Building Stateful Interactive Segmentation with SegVol and GRU Corrector", a novel framework that reformulates interactive segmentation as a sequential decision-making process. Our method introduces: (1) a GRU-based temporal module to model interaction history, enabling dynamic refinement; (2) uncertainty-driven region adaptation to focus corrections on error-prone areas; and (3) a two-stage dynamic loss framework combining global shape consistency with local boundary precision. On 5\% validation data, our framework achieves progressive DSC improvement from 0.661 (single-box prompt) to 0.671 after three refinements, showing 1.5\% absolute gain with diminishing returns in later interactions.

Submission Number: 14