Accurate segmentation of high-throughput platelet cell images based on the YOLOv11 model
10.13303/j.cjbt.issn.1004-549x.2026.08.006
- VernacularTitle:基于YOLOv11模型实现高通量血小板细胞图像精准分割
- Author:
Siyuan ZHANG
1
;
Yan MA
1
Author Information
1. Wuhan Blood Center, Wuhan 430030, China
- Publication Type:Journal Article
- Keywords:
platelets;
optical super-resolution imaging;
cell segmentation;
deep learning
- From:
Chinese Journal of Blood Transfusion
2026;39(8):1033-1038
- CountryChina
- Language:Chinese
-
Abstract:
Objective: To develop an automated platelet instance segmentation method for high-throughput structured illumination microscopy (SIM) data, with the goal of improving segmentation accuracy in wide-field bright-field images. Methods: Bright-field images were used as input to build a YOLOv11-based unified detection-and-segmentation network, trained and inferred with a standardized annotation format. The network consists of a backbone for feature extraction, an improved PAFPN neck for multi-scale feature fusion, and a joint detection/segmentation prediction head, enhancing representation of small targets and crowded scenes. Training was initialized with pretrained weights, using an input resolution of 1 024×1 024 for 200 epochs with a batch size of 4 and automatic mixed precision enabled. Pixels outside the 1st and 99th grayscale percentiles were clipped and linearly mapped to 8-bit intensity values. Data augmentation included flipping, rotation, translation, and scaling. The loss function combined binary cross-entropy loss and Dice loss. Results: In five-fold cross-validation on the gastric cancer dataset, YOLOv11-m achieved balanced precision and recall, with an mAP50-95 of 0.848. Compared with Mask R-CNN, YOLOv8-m, YOLOv9-c, and YOLO26-m, it used a smaller model size and achieved a total processing time of 14.85 ms per image. When directly transferred to validation sets of cholangiocarcinoma, hepatocellular carcinoma, liver cirrhosis, and ovarian cancer, the model maintained stable detection and segmentation performance, indicating good generalization across data sources and imaging variations. Conclusion: The proposed method enables high-throughput, robust platelet instance segmentation and effectively handles challenging bright-field scenarios involving small targets, dense distributions, and local overlaps. By balancing segmentation accuracy and inference efficiency, it serves as a reliable preprocessing component in automated platelet image-analysis pipelines.
