检索历史 清除

目的:研发胎儿脊髓圆锥末端尾侧椎体骨化中心智能计数模型并评估其应用效果。方法:回顾性选取2021年1月至2022年10月深圳市妇幼保健院及珠海市人民医院采集的3 000帧胎儿脊柱骶尾段正中矢状面声像图，其中2 800张图像进行椎体骨化中心与脊髓圆锥的人工精细标注与训练，采用Yolov8算法搭建实例分割模型进行分割训练并通过后处理的方式进行椎体骨化中心拟合与自动计数。另外200张图像分别由人工智能(AI)模型、中级医师(D1)以及初级医师(D2)对胎儿脊髓圆锥末端尾侧椎体骨化中心个数进行计数并计时，由专家对AI模型和D1、D2的计数是否准确进行评估。比较AI模型、D1和D2的准确率以及耗时差异。结果:经过专家评估，AI圆锥分割拟合与计数的准确率达95.00%(190/200)，与D1的准确率(94.50%，189/200)差异无统计学意义( P＝0.823)，但高于D2 88.50%(177/200)，差异有统计学意义( P＝0.012)。D1、D2与AI计数耗时分别为5.00(4.25，6.00)s、7.00(7.00，8.00)s、0.09(0.08，0.10)s，AI明显少于D1、D2，差异有统计学意义(均 P<0.001)。 结论:经过训练的AI模型可以高效、准确地完成圆锥末端的椎体骨化中心计数，相当于中级医师的水平，该模型有望进一步应用于胎儿脊柱裂筛查，提高产前超声筛查的自动化与智能化水平。

Objective:To develop and test the intelligent model for automatically counting the number of vertebral ossification centers below the end of conus medullaris.Methods:From January 2021 to October 2022, 3 000 ultrasound images of the sacrococcygeal spinal middle sagittal plane were retrospectively selected from Shenzhen Maternal and Child Healthcare Hospital and Zhuhai People′s Hospital. The vertebral ossification center and spinal conus medullaris were artificially fine-marked in 2, 800 images for segmentation training. Yolov8 algorithm was used to build the segmentation model for segmentation training, and the fitting and automatic counting of vertebral ossification centers were carried out by post-processing. In the other 200 planes, the counting was performed by the artificial intelligence (AI) model, attending physician (D1), and junior physician (D2), and the accuracy of their performance were evaluated by a specialist physician. The accuracy and the time spent between D1, D2, and AI were compared.Results:The accuracy of AI model segmentation fitting and counting reached 95.00% (190/200) by the specialist physician evaluation, which was almost equal to 94.50%(189/200) by D1( P=0.823) and higher than that of 88.50% by D2(177/200)( P=0.012). The counting time spent for D1, D2, and AI model were 5.00 (4.25, 6.00)s, 7.00 (7.00, 8.00)s, 0.09 (0.08, 0.10)s, respectively, showing that the time spent by AI model was significantly shorter than that of doctors(all P<0.001). Conclusions:The trained artificial intelligence model can efficiently and accurately complete the vertebral ossification center counting below the end of conus medullaris, equivalent to the level of attending physicians. This study is expected to be further applied in the screening of fetal spina bifida and improve the automation and intelligence level of prenatal ultrasound screening.