基于改进DeepLabV3+的囊型肝包虫病超声图像分割算法

Ultrasound image segmentation algorithm for hepatic cystic echinococcosis based on improved DeepLabV3+

目的:将基于改进DeepLabV3+的图像语义分割算法应用到囊型肝包虫病超声图像处理中,实现肝包虫病病灶的自动分割与检测,提高临床诊断效率。方法:本研究采用了DeepLabV3+图像语义分割网络为基础方法,并对其进行了以下改进。首先,为解决DeepLabV3+图像语义分割方法计算复杂度高,内存消耗大,难以在计算能力有限的嵌入式平台上部署,在提取图像特征信息时难以充分利用多尺度信息等问题,以MobileNetV2替换模型的原主干网络Xception,获得轻量级的模型框架。其次,将高效通道注意力应用于底层特征,降低计算复杂度,提高目标边界的清晰度。最后,将Dice Loss引入模型中,缓解模型更关注背景区域,而忽略了包含目标的前景区域等问题。结果:在自建囊型肝包虫病VOC2007数据集5种病灶类型上进行验证,实验结果表明,改进模型的平均交并比和平均像素精度分别达到73.8%和83.5%,能够预测更精细的语义分割结果,有效地优化模型复杂度和分割精度。

Objective To apply the improved DeepLabV3+based image semantic segmentation algorithm to the ultrasound image processing for hepatic cystic echinococcosis,thereby achieving automatic segmentation and detection of hepatic echinococcosis lesions,and improving clinical diagnostic efficiency.Methods DeepLabV3+based image semantic segmentation network was employed as the basic method,and the following improvements were made.To address the issues of high computational complexity,high memory consumption,difficulty in deploying on embedded platforms with limited computing power,and difficulty in fully utilizing multi-scale information when extracting image feature information,the original backbone network Xception of the model was replaced with MobileNetV2 for obtaining a lightweight model framework.Additionally,efficient channel attention was applied to underlying features for reducing computational complexity and improving the clarity of target boundaries;and finally,Dice Loss was introduced into the model to alleviate the problem of the model focusing more on the background area and ignoring the foreground area containing the target.Results Validation was conducted on 5 lesion types in the self-built VOC2007 dataset of hepatic cystic echinococcosis.Experimental results showed that the improved model achieved a mean intersection over union of 73.8 and a mean pixel accuracy of 83.5,indicating that the model can predict more precise semantic segmentation results and effectively optimize model complexity and segmentation accuracy.