Cell Instance Segmentation with Large Vision Model for Circulating Aberrant Cells Identification through Fluorescence In Situ Hybridization Images

Zhenyu REN, Huajia Wang, Yinglan Kuang, Xing Lu

Keywords: Large Vision Model, Cell Segmentation, FISH

Abstract: Liquid biopsy presents a promising non-invasive or minimally invasive approach for early detection of lung cancer. By quantifying 4-color fluorescence in situ hybridization (FISH) signals, circulating genetically abnormal cells (CACs) can potentially be identified with high stability, sensitivity, and specificity. However, precise segmentation of cells is a prerequisite for accurate signal counting. In supervised learning, deep learning models have shown excellent instance segmentation capabilities but require substantial labeled data. To overcome this limitation, we propose a novel zero-shot learning framework for cell instance segmentation. Specifically, we first leverage the Watershed algorithm to generate segmentation proposals and prompts for a large generative vision model, the Segment Anything Model (SAM). We then filter the model outputs based on prior knowledge to obtain pseudo labels for training a specialized instance segmentation model. Our approach eliminates the need for manually labeled data. We demonstrate its effectiveness by segmenting cell images from liquid biopsy and comparing performance against generalized cell segmentation methods(Cellpose). This zero-shot learning paradigm could expand the applicability of vision models to specialized medical imaging applications without costly labeling.

Submission Number: 23