基于DenseASPP模型的超声图像分割

利用超声图像获取胎儿的各项生物指标,对诊断胎儿发育过程中的异常有重要作用.当前主要依靠医生对超声图像的手动测量来确定这些指标.然而,医师手动测量不仅具有主观性,而且在重复作业下效率低下.针对以上问题,提出一种基于DenseASPP模型的超声图像分割改进算法,以辅助医生完成对胎儿各项生物指标的测量.在DenseASPP模型中,首先利用普通卷积预先提取原始图像的特征得到预特征图,再以扩张卷积及金字塔池化结构为基础将前层所有扩张卷积的输出特征图与预特征图拼接在一起传输到下一层扩张卷积以获得更大感受野的多尺度特征图,最终将所有特征合并后通过Attention机制获得相关联的特征,再利用sigmoid函数获取分割结果.分别使用胎儿的头臀径,头围,腹围三个部位的超声图像作为数据集对本文提出的DenseASPP方法进行了评估.实验结果表明,DenseASPP方法优于其他当前常见的分割方法,取得了更好的性能.

Ultrasound image segmentation based on DenseASPP model

Obtaining fetal biological indicators from ultrasound images plays a significant role in diagnosing fetal abnormality. However, manual measurement by physicians is not only subjective, but also leads to inefficiency under repeated operations. To solve the above problems, we propose an improved ultrasound image segmentation algorithm based on DenseASPP model to assist in the measurement of fetal indicators. According to the atrous convolution and structure of Atrous Spatial Pyramid Pooling, the authors firstly extract the pre feature maps of the original image by the ordinary convolution, then the output of each atrous layer is concatenated with the input feature map and all the outputs from lower layers, and the concatenated feature map is fed into the following layer. Finally, all the features are merged to obtain the relevant features through the Attention mechanism, and sigmoid function is used to obtain the segmentation results. We evaluate the method using ultrasound images of fetal head and hip diameters, head circumference, and abdominal circumference as data sets. The experimental results show that this method is superior to other advanced segmentation methods and has better performance.