基于SE-SWTNet的孕妇引产超声图像的关键点自动检测研究

Automated Key Point Detection for Labor Induction Ultrasound Imaging in Pregnant Women Using SE-SWTNet

摘要: 为了降低引产风险并提高妊娠临床诊断的准确性，需要识别产前超声检查中的几个关键参数。收集妊娠第38周至42周的120幅宫颈超声图像，对数据进行标注和预处理。以传统swin transformer (Shifted Window Transformer)网络为基线网络，引入转置卷积和最后的输出特征层，能直接输出关键点参数；同时加入正则化函数和SE (Squeeze and Excitation)注意力机制，减少过拟合风险，提高模型的泛化能力。将SE-SWTNet (Squeeze and Excitation Shifted Window Transformer)与现有的4种算法进行对比。结果显示，SE-SWTNet能较为精确地捕捉宫颈超声图像中的关键点参数。在测试集上，SE-SWTNet的平均半径误差为1.11 mm，并分别在4 mm、5 mm和6 mm误差范围内达到了93.5%、96.5%和97.5%的准确率。SE-SWTNet可以实现基于孕妇引产超声图像的高精度关键点自动检测的功能，为临床医生提供更准确的诊断信息，从而降低风险，提高成功率。

Abstract: In order to reduce the risk of induced abortion and improve the accuracy of clinical diagnosis of pregnancy, it is necessary to identify several key parameters in prenatal ultrasound examination. Collect 120 cervical ultrasound images from weeks 38 to 42 of pregnancy, annotate and preprocess the data. Using the traditional Swin transform er (Shifted Window Transformer) network as the baseline network, introducing transposed convolution and the final output feature layer, it can directly output keypoint parameters; Simultaneously incorporating regularization functions and SE (Squeeze and Excitation) attention mechanisms to reduce the risk of overfitting and improve the model’s generalization ability. Compare SE SWTNet (Squeeze and Excitation Shifted Window Transformer) with four existing algorithms. The results showed that SE SWTNet can accurately capture key point parameters in cervical ultrasound images. On the test set, the average radius error of SE SWTNet is 1.11 mm, and it achieved accuracies of 93.5%, 96.5%, and 97.5% within error ranges of 4 mm, 5 mm, and 6 mm, respectively. SE SWTNet can achieve high-precision automatic detection of key points based on ultrasound images of induced abortion in pregnant women, providing more accurate diagnostic information for clinical doctors, thereby reducing risks and improving success rates.