摘要
胎儿头围是产前超声检查中评价胎儿生长发育最重要的生物特征之一,但手工测量耗时费力且存在操作者的误差。对此,根据超声图像中胎儿头部接近椭圆形状的特征,提出头围测量损失函数。在Mask R-CNN的分割分支后,利用Elli Fit算法对分割掩膜进行椭圆拟合,用Ramanujan公式计算拟合椭圆周长作为头围测量值,将头围真实值和测量值的均方误差作为头围测量损失函数加入原损失函数,使模型训练过程与测量任务紧密相关。对190幅胎儿头部超声图像进行测试,Dice系数为96.89%±1.01%,测量误差为(0.33±1.54) mm,平均处理一幅超声图像的时间为0.33 s。与传统手工测量方法或原模型相比,所提出的方法在速度上提高1.13~16.87 s,在精度上提高0.21~1.68 mm。结果表明,改进的Mask R-CNN可以提高医生测量胎儿头围的效率,能够满足临床需求。
Fetal head circumference is one of the most important biological characteristics in prenatal ultrasound evaluation of fetal growth and development. However,manual measurement is time-consuming and laborconsuming and may have errors by the operator. According to the feature of fetal head close to ellipse shape in ultrasound image,the head circumference measurement loss function was proposed in this paper. After the segmentation branch of Mask R-CNN,ElliFit algorithm was used to fit the ellipse of the segmentation mask.Ramanujan formula was used to calculate the fitting ellipse circumference as the measurement value of the head circumference. The mean square error of the real value of the head circumference and the measurement value was added into the original loss function as head circumference measurement loss function to allow the training process of the model to be closely related to the measurement task. By this way the measurement accuracy and speed was improved. One hundred and ninety ultrasound images of fetal head were tested. Dice’s coefficient was 96. 89% ±1. 01%,and the measurement error was(0. 33±1. 54)mm. The average processing time of one ultrasound image was 0. 33 s. Compared with the traditional manual measurement method or the current machine learning methods,the proposed method improved the speed between 1. 13 seconds and 16. 87 seconds,and improved the accuracy between 0. 21 mm and 1. 68 mm. The results showed that the improved Mask R-CNN increased the efficiency of doctors in measuring fetal head circumference,which met the clinical needs.
作者
李宗桂
张俊华
梅礼晔
Li Zonggui;Zhang Junhua;Mei Liye(School of Information,Yunnan University,Kunming 650500,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2021年第1期12-18,共7页
Chinese Journal of Biomedical Engineering
基金
国家自然科学基金(61841112)。
