Centersam: Fully Automatic Prompt for Dense Nucleus Segmentation

Nucleus segmentation is a vitally important task in biomedical image analysis which leads to multiple applications such as cellular behavior study, tumor detection and cancer diagnosis. However, challenges, such as ambiguous boundary for touching or overlapping nuclei often exist. This paper presents a dense nucleus segmentation method, namely CenterSAM combining the advantages from CenterNet and Segment Anything Model (SAM). It allows fully automatic prompting segmentation without prior knowledge enabling accurate and generalizable nucleus segmentation for biomedical images. Comprehensive evaluations of proposed method are performed on three nucleus segment benchmarks. The results highlight CenterSAM significantly out-performs the second best method by 5.3% on Dice Similarity Coefficient (DSC) in dense nucleus scenarios, meanwhile achieves competitive results on the sparse nucleus segmentation task. The code has been made publicly available
¹
.