Prompt Group-Aware Training for Robust Text-Guided Nuclei Segmentation

Researchers developed a new training method to improve the robustness of AI foundation models like SAM3 for medical image segmentation by reducing sensitivity to prompt variations. The approach groups semantically similar prompts together and uses consistency constraints to ensure more reliable predictions across different prompt formulations.

• →Foundation models for medical image segmentation are highly sensitive to how prompts are formulated, leading to inconsistent results.
• →The new prompt group-aware training framework organizes related prompts into groups sharing the same ground-truth mask.
• →The method combines quality-guided group regularization with logit-level consistency constraints to align predictions.
• →Testing on nuclei segmentation tasks showed 2.16 points improvement in Dice scores across six zero-shot cross-dataset tasks.
• →The approach requires no architectural changes and maintains the same inference process while improving robustness.