摘要
目的构建一种基于超声图像的子宫肌瘤智能识别与病灶精准分割技术, 提高诊断的效率和准确性。方法回顾性纳入2020年11月至2024年10月在深圳市光明区妇幼保健院经阴道超声检查确诊为子宫肌瘤的1 430例患者。超声图像由两位经验丰富的医师进行手动标注, 并由资深专家复核。采用Mask DINO深度学习模型进行病灶分割, 并通过椭圆拟合技术优化分割结果。模型性能通过Dice系数、组内相关系数(ICC)、平均绝对误差(MAE)和测量准确率进行评估。结果在286例测试集中, 模型预测的平均Dice系数为0.992, 显示出极高的分割准确性。模型在病灶识别的平均准确率为0.909, 识别正确样本为241例, 基本正确样本为19例, 错误样本为26例。在长短径测量方面, 模型直接预测的ICC分别为0.871(短径)和0.784(长径), MAE分别为0.436 cm(短径)和0.508 cm(长径)。经椭圆拟合优化后, ICC提升至0.893(短径)和0.866(长径), MAE降低至0.191 cm(短径)和0.274 cm(长径), 测量准确率显著提高。结论本研究构建的基于超声图像的子宫肌瘤智能识别与病灶精准分割技术在病灶分割和测量性能方面表现出色, 能够显著提高诊断的效率和准确性。
Objective:To develop an intelligent recognition and precise segmentation technique using ultrasonic images,and to enhance diagnostic efficiency and accuracy.Methods:A total of 1,430 patients diagnosed with uterine fibroids through transvaginal ultrasonography at the Maternal and Child Health Hospital of Guangming from November 2020 to October 2024 were retrospectively included.Ultrasonic images were manually annotated by two experienced physicians and reviewed by a senior expert.The Mask DINO deep learning model was used for lesion segmentation,and the segmentation results were optimized using ellipse fitting technology.Model performance was evaluated using the Dice coefficient,intraclass correlation coefficient(ICC),mean absolute error(MAE),and measurement accuracy.Results:In the test set of 286 cases,the average Dice coefficient of model prediction was 0.992,indicating extremely high segmentation accuracy.The average accuracy of lesion identification by the model was 0.909,with 241 correctly identified samples,19 basically correct samples,and 26 incorrect samples.In terms of long and short axis measurements,the ICC of the model's direct predictions were 0.871(short axis)and 0.784(long axis),with MAE of 0.436 cm(short axis)and 0.508 cm(long axis).After optimization with ellipse fitting,the ICC increased to 0.893(short axis)and 0.866(long axis),and the MAE decreased to 0.191 cm(short axis)and 0.274 cm(long axis),the measurement accuracy improved significantly.Conclusions:The intelligent recognition and precise segmentation technique for uterine fibroids based on ultrasonic images constructed in this study performed excellently in lesion segmentation and measurement,it can significantly improve the efficiency and accuracy of diagnosis.
作者
张燕辉
熊奕
石波
梁晓冰
陈美兰
吴凯
Zhang Yanhui;Xiong Yi;Shi Bo;Liang Xiaobing;Chen Meilan;Wu Kai(Shantou University of Medicine,Shantou 515000,China;Department of Ultrasound,Maternal and Child Health Hospital of Guangming District,Shenzhen 518107,China;Department of Ultrasound,Shenzhen Luohu People's Hospital,Shenzhen 518005,China;*Department of Ultrasound,Shenzhen Guangming District People's Hospital,Shenzhen 518107,China)
出处
《中华超声影像学杂志》
北大核心
2025年第7期602-607,共6页
Chinese Journal of Ultrasonography
基金
广西重点研发计划(桂科AB23026042)。
关键词
子宫肌瘤
超声图像
深度学习
椭圆拟合
自动测量
Uterine fibroids
Ultrasound image
Deep learning
Ellipse fitting
Automatic measurement
