Efficient Cell Segmentation and Classification by Leveraging an Auto Prompt Generator and a New Fine-Tuning Strategy

Cell nucleus instance segmentation is an important task in biomedical image analysis. However, due to complex nuclear shapes, clustering, overlap, and blurred boundaries, it remains a challenging problem. In this study, to address the complexities of cell nucleus segmentation and enable classification, we propose a hybrid model that enhances performance and enables classification by combining a new fine-tuning method of SAM, the HQ-token and global-local feature fusion block of HQ-SAM, and an automatic prompt generator (YOLOv8). As a result, the proposed method achieved an F1 detection score of 85.73 on the PanNuke dataset.