兰州大学学报(医学版)

目的 构建人工智能辅助系统，实现小肠胶囊内镜(SBCE)图像中8类病变的自动检测与分割。方法 整合3个数据集的SBCE图像，使用LabeMe工具多边形标注病变，并转换为YOLO模型兼容的格式，用于神经网络模型的训练、验证与测试，包含13 983张图像和17 911个注释标签。模型性能评估指标包括精确率、灵敏度、50%交并比阈值下的平均精度(mAP⁵⁰)、50%～95%交并比阈值下的平均精度均值(mAP^50～95)、推理速度等指标。结果 共开发出5种不同规模的YOLO v11分割模型(v11n、v11s、v11m、v11l、v11x)。其中，YOLO v11m在保持最高mAP⁵⁰(0.908)的同时，实现了较快的推理速度(208.3帧/s)，因此被选为最佳模型。在外部测试集中，YOLO v11m对8类小肠病变的分割任务取得了0.892的总体mAP⁵⁰，其中息肉和淋巴管扩张的分割精度最高，mAP^50～95分别为0.723和0.707，而出血类别的mAP^50～95最低，仅为0.409。结论 基于YOLO v11神经网络开发的SBCE图像分割模型YOLO v11m具有良好的识别性能，能够自动实现多类别小肠病变的精准定位、分类，并准确勾勒病变的像素级轮廓。

Objective An artificial intelligence-assisted system was developed for the automatic detection and segmentation of eight types of small bowel lesions in small bowel capsule endoscopy(SBCE) images. Methods SBCE images were collected from three datasets, and lesion annotations performed using the LabeMe tool with polygonal segmentation, later converted into a You Only Look Once(YOLO)-compatible format for training, validation and testing of the neural network model. The dataset comprised 13,983 images with 17 911 annotated labels. Model performance was evaluated using precision, sensitivity, mean average precision at an intersection-over-union(IoU) threshold of 50%(mAP⁵⁰), mean average precision across IoU thresholds from 50% to 95%(mAP^50～95), and inference speed. Results Five YOLO v11 segmentation models of different scales(v11n, v11s, v11m, v11l, v11x) were developed. Among them, YOLO v11m achieved the highest mAP⁵⁰(0.908) while maintaining a fast inference speed of 208.3 frames per second, making it the optimal model. On the external test set, YOLO v11m attained an overall mAP⁵⁰ of 0.892 for segmenting the eight lesion types. The highest segmentation accuracy was observed for polyps and lymphangiectasia, with mAP^50～95 values of 0.723 and 0.707 respectively, whereas the lowest performance was noted for the bleeding category, with an mAP^50～95 of only 0.409. Conclusion The SBCE segmentation model YOLO v11m was developed based on the YOLO v11 neural network demonstrated strong lesion recognition capabilities, enabling precise multi-class lesion localization, classification and pixel-level contour delineation. These findings highlight its promising potential for clinical applications.