联合超声甲状腺结节分割与分类的多任务方法研究

Multi-task method for segmentation and classification of thyroid nodules combined with ultrasound images

针对超声图像中甲状腺结节多尺度、结节边缘模糊、良恶分类不平衡问题,提出一种联合超声甲状腺结节分割与分类的多任务方法。以全卷积网络作为主干共享网络,将提取到的浅层特征共享给多任务分支网络,在分割网络分支中,先加入深层卷积块,获取分割分支深层特征,再对深层特征进行上采样。本文提出一种改进卷积注意力模块的多尺度卷积注意力模块,将上采样结果与主干共享网络每个特征提取阶段经过带有多尺度卷积注意力模块跳跃连接后的特征张量进行拼接,减少结节边缘模糊问题,提高分割性能。同时将多尺度卷积注意力模块融入到分类分支中,优化分类性能。实验结果表明:本文所提多任务方法能有效提升分割和分类的精度,较单任务深度学习网络具有更优的分割与分类性能,能有效处理甲状腺结节多尺度、结节边缘模糊的问题,降低良恶分类不平衡带来的影响。

Aiming at the problems of multi-scale thyroid nodules,blurred nodule edges,and unbalanced classification of benign and malignant thyroid nodules in ultrasound images,this paper proposes a multi-task method for segmentation and classification of thyroid nodules combined with ultrasound.The fully convolutional network is used as the backbone sharing network,and the extracted shallow features are shared to the multi-task branch network.In the branch segmentation networks,deep convolution blocks are added to obtain the deep features of the segmented branches,and then the deep features are up-sampled.An improved multi-scale convolutional attention module is proposed,which combines the up-sampling results with the feature tensor of each feature extraction stage of trunk sharing network after jumping connection with multi-scale convolution attention module,so as to reduce the fuzzy problem of nodule edge blurs and improve the segmentation performance.At the same time,a multi-scale convolutional attention module is integrated into the classification branch to optimize the classification performance.The experimental results show that the multi-task method proposed in this paper can effectively improve the accuracy of segmentation and classification,having better segmentation and classification performance than single-task deep learning network.It can effectively deal with the problem of multi-scale thyroid nodules and blurred nodule edges,and reduce the impact brought by unbalanced classification of benign and malignant.