基于MDL-U2-Net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
-
摘要: 准确地分割超声图像中盆底区域,是实现盆底疾病计算机辅助诊断的重要环节.针对盆底形态复杂、边界模糊、分割算法参数量庞大以及参数测量精度有限等问题,搭建了一种轻量级语义分割网络MDL-U2-Net并提出修补算法AC-F.首先,对基准U2-Net进行结构优化和通道数调整,以有效地降低模型参数量;其次,融入复合损失函数以缓解训练损失波动并提升边界保持能力,提高网络对模糊边界的分割准确性;之后,提出深度非对称多尺度混洗卷积模块,以捕获特征空间采样的位置偏移信息,弥补轻量网络感受野不足和特征提取能力较弱的缺陷,提高网络对盆底复杂形态的建模能力;最后,采用修补算法对分割盆底进行精细化填补,以提高盆底完整性和参数测量的精度.在自制数据集上的实验结果表明,MDL-U2-Net对盆底分割的Jaccard,Recall和HD95指标分别达到91.226%,93.589%和1.074,与基准U2-Net相比,模型参数量缩减了94.37个百分点;此外,经AC-F算法处理后的区域面积测量百分误差降至1.25%,ICC达到0.998且有95%(76/80)的数据在95% LoA内,能够实现轻量级分割和精确参数测量.Abstract: Precise segmentation of the pelvic floor area in ultrasound images represents a critical component in the realization of computer-aided diagnosis for pelvic floor diseases. Aiming at the problems of complex pelvic floor shape, blurred boundary, huge number of segmentation algorithm parameters and limited parameter measurement accuracy, a minimizing semantic segmentation network MDL-U2-Net and patching algorithm AC-F were built forward in this research. First, the structure optimization and channel number adjustment of the benchmark U2-Net are carried out to reduce the amount of model parameters; secondly, the composite loss function is integrated to alleviate training loss fluctuations and improve the boundary retention ability, improving the segmentation accuracy of the network for blurred boundaries; following this, a Deep Asymmetric Multi-scale Shuffled Convolutional Module is introduced to capture positional offset information during feature space sampling. This module compensates for the limited receptive fields and suboptimal feature extraction capabilities inherent in lightweight networks, ultimately enhancing the network’s capacity to model the intricate structures of the pelvic floor; finally, the patching algorithm is used to refine and fill the segmented pelvic floor to improve the integrity and accuracy of parameter measurement The experimental results on the self-made dataset show that the Jaccard, Recall and HD95 indexes of MDL-U2-Net for pelvic floor segmentation reach 91.226%, 93.589% and 1.074, respectively. Compared with the benchmark U2-Net, the model parameters are reduced by 94.37 percentage points. In addition, error of the regional area measurement processed by AC-F algorithm is reduced to 1.25%, the ICC reaches 0.998 and 95%(76/80) of the data is within the 95% LoA, enabling lightweight segmentation and accurate parameter measurement.