高级检索
刘孝保, 甘博敏, 姚廷强, 申吉泓. 基于MDL-U2-Net网络的盆底超声图像轻量级分割及参数测量[J]. 计算机辅助设计与图形学学报. DOI: 10.3724/SP.J.1089.2023-00063
引用本文: 刘孝保, 甘博敏, 姚廷强, 申吉泓. 基于MDL-U2-Net网络的盆底超声图像轻量级分割及参数测量[J]. 计算机辅助设计与图形学学报. DOI: 10.3724/SP.J.1089.2023-00063
XiaoBao LIU, BoMin GAN, TingQiang YAO, JiHong SHEN. Minimizing segmentation and parameter measurement of pelvic floor ultrasound images based on MDL-U2-Net network[J]. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/SP.J.1089.2023-00063
Citation: XiaoBao LIU, BoMin GAN, TingQiang YAO, JiHong SHEN. Minimizing segmentation and parameter measurement of pelvic floor ultrasound images based on MDL-U2-Net network[J]. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/SP.J.1089.2023-00063

基于MDL-U2-Net网络的盆底超声图像轻量级分割及参数测量

Minimizing segmentation and parameter measurement of pelvic floor ultrasound images based on MDL-U2-Net network

  • 摘要: 针对超声图像中盆底形态多变、边缘模糊以及现有终端设备计算能力有限的问题, 提出一种轻量级语义分割算法DML-U2-Net, 对于盆底参数测量精度不高的问题, 提出AC-F修补方法进行精细化处理. 首先, 对基准U2-Net进行结构优化和通道数调整, 有效降低了网络参数量, 并提出一种基于Log-SD损失的层间加权混合损失函数来缓解网络训练过程中的损失波动, 以提高网络的边界保持能力, 解决盆底边界模糊导致边缘分割精度较低的问题; 其次, 引入DAMS卷积模块替换网络中部分单一卷积层, 以弥补小型网络感受野不足和特征提取能力较弱的缺陷, 提升网络对形态多变盆底目标的适应性. 最后, 利用AC-F修补方法对盆底区域和轮廓进行精细化处理, 以提高盆底完整性和参数测量的精度. 实验结果表明, DML-U2-Net对盆底超声分割Jaccard、Recall和HD95指标分别达到91.226%、93.589%和1.074, 且模型参数量缩减了94.37%, 降至11.4M; 此外, 经过AC-F处理后的RA测量百分误差更小, 为1.25%, ICC达到0.998且有95%(76/ 80)的数据在95%LoA范围内, 能够实现轻量级分割和精确参数测量.

     

    Abstract: Aiming at the problems of changing pelvic floor shape, blurred edges and limited computing power of existing end point equipment in ultrasonic images, a lightweight semantic segmentation algorithm DML-U2-Net is proposed. For the problem of low pelvic floor parameter measurement accuracy, AC-F is proposed. The repair method is refined. Firstly, the structure optimization and channel number adjustment of the benchmark U2-Net are carried out, which effectively reduces the amount of network parameters, and an interlayer weighted hybrid loss function based on Log-SD loss is proposed to alleviate the loss fluctuation during the network training process, so as to improve the network. The boundary retention ability of the network solves the problem of low edge segmentation accuracy caused by the blurred boundary of the pelvic floor; secondly, the DAMS convolutional module is introduced to replace part of the single convolutional layer in the network to make up for the shortcomings of the small network receptive field and weak feature extraction ability, and improve the network's adaptability to the morphological pelvic floor target. Finally, the area and contour of the pelvic floor are refined by the AC-F repair method to improve the integrity of the pelvic floor and the accuracy of parameter measurement. The experimental results show that the Jaccard, Recall and HD95 indexes of pelvic floor ultrasonic segmentation by DML-U2-Net reach 91.226%, 93.589% and 1.074, respectively, and the model parameters are reduced by 94.37% to 11.4M; in addition, the RA measurement after AC-F treatment The percentage error is smaller, 1.25%, ICC reaches 0.998 and 95% (76/80) of the data is within the 95% LoA range, which can realize lightweight segmentation and accurate parameter measurement.

     

/

返回文章
返回