One-shot cell segmentation via learning memory query: Towards universal solution without active tuning.

Abstract

Cell segmentation, which involves separating individual cells in biomedical images, is essential for disease analysis and drug development research. However, many existing methods are restricted to specific types of images or require constant adjustment, making them time-consuming and labor-intensive. We introduce a new framework called Mimic, which employs a "Query-and-Answer" (Q&A) mechanism to segment cells in a single step. This innovative approach eliminates the need for constant adjustments across different images, significantly reducing labor-intensive tasks. Mimic learns to recognize and segment cells using a few examples as "prompts", allowing this model to adapt to new cell types without additional training. Mimic was tested on 12 public datasets featuring various imaging techniques, cell shapes, sizes, and staining methods. It achieved state-of-the-art performance, surpassing existing generalist cell segmentation models such as Cellpose and Stardist and foundational vision models. Mimic's capability to segment cells without extensive tuning or additional training could greatly enhance the speed and accuracy of quantitative analysis in biological and medical research.