AA-NuSeg: a fully automated and adaptive framework for accurate nuclei segmentation

Abstract

Nuclei segmentation faces challenges due to diverse imaging modalities and cell morphologies. While the Segment Anything Model (SAM) excels in universal tasks, its use in medical nuclei segmentation is constrained by manual prompting and limited adaptability. To overcome this, we propose AA-NuSeg, a fully automated, self-prompted segmentation framework built on SAM. AA-NuSeg introduces three key components: a Cross-Domain Feature Adapter (CDF-Adapter) for domain-specific knowledge integration, an Image-driven Prompt Generation Network (IPG-Net) that generates automatic semantic prompts, and a Semantic-Instance Refinement Decoder (SIR- Decoder) for precise instance segmentation. Evaluations confirm AA-NuSeg outperforms existing methods, offering superior accuracy and robust domain generalization.

Ähnliche Arbeiten