多阶U-Net甲状腺超声图像自动分割方法

Multi-Stage U-Net Automatic Segmentation of Thyroid Ultrasound Images

甲状腺超声图像广泛应用于甲状腺相关疾病的诊断。针对甲状腺超声图像对比度低、边缘模糊以及散斑噪声严重等问题,提出一种基于多阶U-Net的深度卷积网络模型,用于实现甲状腺腺体与甲状腺结节的自动分割。该模型以U-Net为基本网络框架,通过不断进阶的特征融合,以实现图像边缘的信息提取。同时,在模型中使用了一种多尺度残差卷积模块以进一步提升分割精度。对比实验结果表明,该模型相较于其他方法能够获得更好的分割结果,具有一定的临床应用价值。

Thyroid ultrasonography is widely used in the diagnosis of thyroid diseases. To solve the problems of low contrast, blurred edges and serious speckle noise in thyroid ultrasound images, a deep convolutional network model based on multi-stage U-Net is proposed to achieve automatic segmentation of thyroid glands and thyroid nodules. Using U-Net as the basic network framework, this model realizes the deep information extraction of image edge through continuously advanced feature fusion. Meanwhile, a multi-scale residual convolution module is used in the model to further improve the segmentation accuracy. The comparative experimental results show that this model can obtain better segmentation results compared with other methods, which has certain clinical application value.