Reproduce “Unleashing the Power of Prompt-driven Nucleus Instance Segmentation” on OpenResearch

About this paper

Zhongyi Shui, Yunlong Zhang, Kai Yao, Chenglu Zhu, Sunyi Zheng, Jingxiong Li, Honglin Li, Yuxuan Sun and 2 more

Researchers from Zhejiang University, University of Liverpool, and Westlake University introduced PromptNucSeg, a prompt-driven framework for nucleus instance segmentation that leverages a fine-tuned Segment Anything Model. The method achieved new state-of-the-art performance on PanNuke, Kumar, and CPM-17 datasets, while offering improved computational efficiency and removing the need for complex post-processing.

Key insights

• A prompt-driven, end-to-end framework can achieve superior nucleus instance segmentation without requiring the complex and sensitive post-processing steps common in traditional bottom-up methods.
• Effective adaptation of foundational models like SAM for specific medical imaging tasks requires fine-tuning not just the encoder but also adapting the prompt-driven mechanism to the target domain's instance-level requirements.
• Integrating negative prompts during training is crucial for the segmentor to effectively utilize them during inference, preventing 'catastrophic forgetting' and significantly improving the separation of overlapping instances.

Summary by alphaXiv.