摘要
基于超声影像对甲状腺结节进行精准分割,可以得到病变区域的生理参数信息,从而对甲状腺结节的早期筛查和诊断定性.为实现甲状腺结节的精准分割,提出了一种基于Transformer编码的多层次特征融合网络.针对不同患者的甲状腺结节大小和其在甲状腺超声图像中所处的位置均存在较大差异的特点,该模型以Transformer结构作为特征提取器,使各层次特征的计算都在更大、更灵活的感受野上进行;以CNN作为特征解码器,对编码器所获得的特征进行重构,并降低算法复杂度.编码器与解码器之间采用长距离跳跃连接的方式相连.利用局部-全局策略学习甲状腺超声图像中浅层的全局特征和深层的局部特征.此外,通过将模型中的多头注意力机制改进为残差轴向注意力机制,学习到了甲状腺结节中更多的方向纹理特征.实验数据来源于天津医科大学总医院超声影像科,通过对3828例样本采用旋转变换、翻转变换和随机裁剪3种数据增强方法,得到15312例甲状腺超声图像.经过多轮迭代训练,得到测试集样本上的Dice系数为92.2%,交并比为85.5%.相同数据集上的对比实验表明:相对于全卷积神经网络,该算法在Dice系数上提升了5%~8%,在交并比上提升了7%~13%,模型参数量平均降低了5.67×10^(6),精准地实现了甲状腺结节的全自动分割,降低了模型复杂度,具有一定的临床价值.
The accurate segmentation of thyroid nodules based on ultrasound images can obtain physiological parameter information of the lesion area to screen and diagnose thyroid nodules in the early stage.To achieve an accurate segmentation of thyroid nodules,a multilevel feature fusion network based on Transformer coding was proposed.Given the large differences in the sizes of the thyroid nodules and their positions in the thyroid ultrasound images of different patients,the Transformer structure was used as the feature extractor to perform the calculation of features at each level on a larger and more flexible receptive field.The convolutional neural network was used as the feature decoder to reconstruct the features obtained by the encoder and reduce the complexity of the algorithm.The encoder and the decoder were connected by a skip connection.A local-global strategy was used to learn the local and global features in the thyroid ultrasound images.Additionally,more directional texture features in thyroid nodules were learned by improving the multi-head attention mechanism in the model to the residual axial-attention mechanism.The experimental data came from the Department of Ultrasound Imaging of the General Hospital of Tianjin Medical University.Using three data augmentation methods,including the rotation transformation,flip transformation,and random cropping,on 3 828 samples,15 312 thyroid ultrasound images were obtained. After multiple rounds of iterative training,the Dice coefficient on the test set was 92.2%,and the intersection ratio was 85.5%. Comparative experi-ments on the same data set showed that compared with the full convolution neural network,the algorithm improved the Dice coefficient by 5% to 8%,increased the intersection ratio by 7% to 13%,and reduced the average algorithm complexity by 5.67×10^(6),realizing the automatic segmentation of the thyroid nodules accurately and reducing the model complexity. Results show that it has a certain clinical value.
作者
曹玉珍
郑洁
余辉
王飞
张杰
Cao Yuzhen;Zheng Jie;Yu Hui;Wang Fei;Zhang Jie(School of Precision Instruments and Optoelectronics Engineering,Tianjin University,Tianjin 300072,China;Academy of Medical Engineering and Translational Medicine,Tianjin University,Tianjin 300000,China;Department of Ultrasound Imaging,Tianjin Medical University General Hospital,Tianjin 300052,China)
出处
《天津大学学报（自然科学与工程技术版）》
EI
CAS
CSCD
北大核心
2022年第7期674-681,共8页
Journal of Tianjin University：Science and Technology
基金
国家重点研发计划资助项目(2019YFC0119402)
天津科技重大专项与工程资助项目(18ZXZNSY00240).
