Ultrasonic image segmentation of thyroid nodules-relevant multi-scale feature based h-shape network

目的 准确定位超声甲状腺结节对甲状腺癌早期诊断具有重要意义，但患者结节大小、形状以及位置的不确定性极大影响了结节分割的准确率和模型的泛化能力。为了提高超声甲状腺结节分割的精度，增强泛化性能并降低模型的参数量，辅助医生诊断疾病，减少误诊，提出一种面向甲状腺结节超声图像分割的多尺度特征融合“h”形网络。方法 首先提出一种网络框架，形状与字母h相似，由一个编码器和两个解码器组成，引入深度可分离卷积缩小网络尺寸。编码器用于提取图像特征，且构建增强下采样模块来减少下采样时造成的信息损失，增强解码器特征提取的能力。第1个解码器负责获取图像的初步分割信息；第2个解码器通过融合第1个解码器预先学习到的信息来增强结节的特征表达，提升分割精度，并设计了融合卷积池化金字塔实现多尺度特征融合，增强模型的泛化能力。结果 该网络在内部数据集上的Dice相似系数（Dice similarity coefficients，DSC）、豪斯多夫距离（Hausdorff distance，HD）、灵敏度（sensitivity，SEN）和特异度（specificity，SPE）分别为0.872 1、0.935 6、0.879 7和0.997 3，在公开数据集DDTI （digital database thyroid image）上，DSC和SPE分别为0.758 0和0.977 3，在数据集TN3K （thyroid nodule3 thousand）上的重要指标DSC和HD分别为0.781 5和4.472 6，皆优于其他模型。结论 该网络模型以较低的参数量提升了甲状腺超声图像结节的分割效果，增强了泛化性能。;Objective Early diagnosis of cancer-beneficial lesions ultra-sound nodules are required to be located accurately. Ultra-sound imaging technique is potential for the diseases and it cost-effective simplified a certain extent. Thyroid reporting data system（TI-RADS）is focused on benign malignant nodules-relevant evaluation recently. The probability nodule will much more distorted when level higher. first step ultrasound oriented segment nodule. At present，the commonly-used segmentation method manual segmentation，which still labor intensive experience behavioral. Computer technology-based medical realizing automatic ultrasonic speed accuracy can improved. Current deep learning has its potentials several visual recognition tasks in recent years. Compared traditional contour-shape region based methods，deep technology preferred improve tasks. Fully convolutional neural network（FCN）and network（CNN）based multiple models achieve specific However，the speckle noise images uncertainty size，shape location patient’ s have affected greatly. Method First，the h-shape network framework proposed terms an encoder two decoders. shape similar letter“h”，and depth separable convolution introduced shrink size. second each layer replaced by lower number parameters model. used extract image features，and enhanced down-sampling module constructed alleviate down-sampling-led information loss. composed connection maximum pool average pool，batch normalization pool，which enhance feature extraction capability decoder. decoder supporting preliminary image，and expression nodules，and improved via decoder-related fusion learned information. Finally，the pyramid pooling designed as well，in which atrous spatial integrated together realize multi-scale while size generalization ability model optimized. four sorts blocks operated through them，and final prediction result generated after concat operation. Three kinds datasets provided verify well，which consists 3 622 images-within internal dataset，637 images-involved digital dataset，and 493 images-included TN3K public dataset. dataset divided into training set，validation set，and test set ratio 8∶1∶1 train Due small samples linked DDTI dataset，partitioning prone over fitted，and weight straightforward. experiment built Pytorch Nvidia RTX 2080 TI Using Adam optimizer， initial rate 0. 000 1. There 200 rounds training，and half every 20 rounds. batch 8. To perform well basis stable overall segmentation， DiceBCELoss regarded loss function，which combine BCE function with Dice function. results analyzed quantitative coefficient（DSC），Hausdorff distance（HD），sensitivity（SEN），and specificity（SPE）. Result validate method，comparative analysis carried out，which comparison AttentionUNet，marker-guided U-Net（MG-UNet），fully dense dilated Net（FCdDN），DeepLab V3+，segmentation network（segNet）and context network（CE_NET）. For dataset， DSC， HD， SEN SPE indexes reached 872 1， 935 6， 879 7 997 each. DSC 15. 53% than worst model，and 2% best one；The HD 2. 583 6 decreased 034 1 model；The increased 32% 17% model，which 7. 57% 9. 96% higher dataset，the 758 0 977 each，which 83% 25% model，1. 02% 71% 781 5 4. 472 obtained dataset，which 27% 634 performance. Furthermore，a series ablation experiments conducted demonstrate effectiveness different steps algorithm. Conclusion This computing cost optimized further.