基于深度学习的肺癌患者颈部淋巴结良恶性辅助超声诊断

ASSISTED ULTRASOUND DIAGNOSIS OF BENIGN AND MALIGNANT CERVICAL LYMPH NODES IN LUNG CANCER PATIENTS BASED ON DEEP LEARNING

深度学习技术辅助超声影像诊断可以提高检测的精度和效率.提出一种用于超声图像分割的改进UNet卷积网络.该网络将噪声激励函数NHReLU和NHSeLU代替ReLU噪声激励函数,对成本函数增加权重参数.通过在两个尺度上预测,很好地处理了超声图像中标注区域尺寸变化的问题,提高对淋巴结超声图像分割效果.使用VGG、ResNet和DenseNet等网络预测淋巴结病灶区域的良恶性.实验表明,分割网络性能优异,Dice系数达到0.90,模型能够很好防止过拟合.在小样本下预测良恶性各指标都得到提高,为深度学习技术应用于超声图像检测提供了新方法.

Applying deep learning to assist ultrasound image diagnosis can improve the accuracy and efficiency of detection.Therefore,this paper proposes a modified UNet convolutional neural network for ultrasound image segment.The noisy activation functions that called Rectified Linear Unit(ReLU)was replaced by Noisy Hard Rectified Linear Unit(NHReLU)and Noisy Hard Scaled Exponential Linear Unit(NHSeLU)in this network.It also added weight parameter to cost function.By predicting on two scales,the problem of dimension change of labeled area in ultrasound image was well handled,and the segmentation effect of lymph node ultrasound image was improved.We used VGG,ResNet and DenseNet to predict the benign and malignant lesions in lymph node lesions.Experiments show that the performance of the segmentation network is excellent,and the Dice coefficient reaches 0.90.The model can prevent over-fitting,and the prediction of benign and malignant indicators are improved under small samples.It provides a new method for the application of deep learning technology in ultrasound image detection.