README.md

Introduction

This is the official github of the paper: CenterSAM: Fully Automatic Prompt for Dense Nucleus Segmentation.

The paper has been accepted by ISBI 2024 (International Symposium on Biomedical Imaging 2024), doi TBD.

CenterSAM is a fully automatic prompting segmentation approach which enabling accurate and generalizable nucleus segmentation for biomedical images. The Figure1 shows the overall archtecture of CenterSAM.

Figure1: Architecture of CenterSAM

Evaluation

The method have been evaluated on three different sized medical image datasets:

• 2018 Data Science Bowl - Represents for "small" dataset

  • Number of Images: 51
  • Annotations: 32,217

• MoNuSeg - Represents for "medium" dataset

  • Number of Images: 670
  • Annotations: 29,461

• TissueNet - Represents for "large" dataset

  • Number of Images: 6990
  • Annotations: ~1.2 Million

Below Table shows quantitative results of comparison against State-Of-The-Art (SOTA) methods. The best results are highlighted in bold

2018 Data Science Bowl Method DSC(%)↑ mIoU(%)↑ UNet++ 91.10 83.70 Deeplabv3+ [paper] [github] 88.80 83.70 SSFormer-S [paper] [github] 92.50 86.50 DuAT [paper] [github] 92.60 87.00 CenterSAM 92.20 86.60	MoNuSeg Method DSC(%)↑ AJI(%)↑ UNet 74.56 60.22 UNet++ [paper] [github] 80.33 67.30 MAE [paper] [github] 73.68 58.62 MDM [paper] 81.01 68.25 CenterSAM 81.95 68.75	TissueNet Method DSC(%)↑ SEG(%)↑ Detectron2 [github] 75.50 78.00 Cellulus [paper] [github] 64.10 52.40 StarDist [paper] [github] 59.40 38.20 Mesmer [paper] [github] 83.40 77.20 CenterSAM 88.70 79.50

Visualization

Tutorial

• Installation
• Dataset Prepare
• Quick Start