分子影像学杂志2025，Vol.48Issue(9)：1057-1063,7.DOI:10.12122/j.issn.1674-4500.2025.09.01
基于YOLO v11的X线气腹检测模型研究
Research on a YOLO v11-based model for pneumoperitoneum detection in X-ray images
摘要
Abstract
Objective To explore the feasibility of applying the deep learning object detection algorithm YOLO v11 to the detection of pneumoperitoneum in upright chest and abdominal X-ray images.Methods Upright anteroposterior chest and abdominal X-ray images were retrospectively collected from five medical centers from January 2014 to December 2023.After screening,a total of 1844 X-ray images(860 chest images and 984 abdominal images)were included in this study.Among them,1728 images from four centers were allocated for model training and internal validation,while 116 images from one center served as an independent external test set.Two YOLO v11 detection models were built on the PyTorch framework:Subdiaphragmatic Lucent(Linear)Shadow Detection Model(Model 1),which exclusively detects subdiaphragmatic lucent(linear)shadows;Multi-Sign Pneumoperitoneum Detection Model(Model 2),which detects subdiaphragmatic lucent(linear)shadows and five additional pneumoperitoneum signs.Results For Model 1,the mAP@0.5 values in internal and external validation were 0.858 and 0.748,the precision values were 0.846 and 0.790,the recall values were 0.791 and 0.709,respectively,and the detection time per image frame was 15.8 ms.For Model 2,the categories with good performance in internal and external validation were subdiaphragmatic lucent(linear)shadows(mAP@0.5:0.859 in internal validation and 0.743 in external validation)and external contours of parenchymal organs against a background of gaseous lucency(mAP@0.5:0.776 in internal validation and 0.766 in external validation).The mAP@0.5 of external contours of overlapping parenchymal organs with low contrast was 0.482 in the external validation set(0.431 in the internal validation set),while the detection efficiency for other signs was low,and the detection time per image frame of Model 2 was 15.5 ms.Conclusion The YOLO v11-based X-ray pneumoperitoneum detection model demonstrates high detection efficiency for subdiaphragmatic lucent(linear)shadows and external contours of parenchymal organs against a background of gaseous lucency.It helps improve the detection rate of pneumoperitoneum signs,promptly indicates the risk of gastrointestinal perforation in patients,and optimizes clinical workflows.关键词
消化道穿孔/气腹/深度学习/目标检测Key words
gastrointestinal perforation/pneumoperitoneum/deep learning/object detection引用本文复制引用
林炯彬,曾伟雄,胡培铅,邓传颂,詹汉钦,邢培华,何景萍,汪思娜,陈卫国..基于YOLO v11的X线气腹检测模型研究[J].分子影像学杂志,2025,48(9):1057-1063,7.基金项目
国家自然科学基金(82171929)Supported by National Natural Science Foundation of China(82171929) （82171929）
