盆底超声智能识别及半自动测量泌尿生殖裂孔的实用性研究

Study on the practicality of the semi-automatic measurement for the urogenital hiatus

目的探讨经会阴盆底超声智能识别及半自动测量与手动测量泌尿生殖裂孔(urogenital hiatus,UH)的一致性。方法将286幅UH最小裂孔三维图像随机分为学习组100幅和测试组186幅。3名盆底超声经验丰富的医生应用离线软件标记学习组,并通过"机器学习算法"获取半自动测量软件。2名盆底超声经验丰富的医生(D1、D2)分别用盆底智能识别及半自动测量软件和手动描迹法对测试组进行测量,测量参数包括UH前后径、左右径、面积、周长、左肛提肌尿道间隙、右肛提肌尿道间隙,分别记录测量耗时,并对耗时和所得UH测值进行统计分析。结果半自动测量耗时:D1为(7.49±1.51)s,D2为(7.52±1.37)s;手动测量耗时:D1为(42.42±11.08)s,D2为(43.45±9.09)s,半自动测量比手动测量明显节省时间(t=-12.09,-13.64;均P=0.00)。两位测量者半自动测量结果与手动测量的Pearson相关系数r分别为0.857~0.985、0.853~0.979,均P<0.01;两位测量者半自动测量与手动测量的ICC分别为0.846~0.985、0.843~0.979;Bland-Altman图显示两种测量方法测值的一致性好。结论盆底超声智能识别及半自动测量的可靠性高,可以简化UH的测量步骤、缩短检查时间,可应用于临床工作。

Objective To determine the consistency of urogenital hiatus (UH) data between the semi-automatic measurement and manual measurement using transperineal pelvic floor ultrasonography. Methods Total of 286 three-dimensional images of minimal UH dimension were obtained. And they were divided into study group (100 images) and test group (186 images) randomly. Three experts traced and created the whole profile of the UH of those images in the study group by MATLAB. Then the semi-automatic software was obtained through machine learning algorithms. In the test group, 6 parameters of UH (including anterioposterior diameter, transverse diameter, circumference, area, left and right levator urethral gap distance) were measured by two experts (D1 and D2) both manually and semi-automatically. The time experts spent on measuring was also recorded and compared. Results The time used for semi-automatic measurement was significantly shorter than that for manual measurement[(7.49±1.51)s vs (42.42±11.08)s,(7.52±1.37)s vs (43.45±9.09)s for D1 and D2, t=-12.09,-13.64, all P=0.00]. The Pearson correlation coefficients between semi-automatic and manual measurements of 6 parameters were 0.857-0.985 (P<0.01), 0.853-0.979 (P<0.01) in D1 and D2, respectively. The interclass correlation coefficients (ICC) of six parameters were ranged from 0.846-0.985 for D1 and 0.843~0.979 for D2(all P<0.01). The Bland Altman plot also showed good agreement between two methods. Conclusions Intellectual recognition and semi-automatic measurement has simplified the process for UH measurement, and it is proved to be a reliable and timesaving method that is practical for clinical use.