Abstract: Thyroid nodule detection plays an important role in medical diagnosis. The traditional machine learning method has many problems, such as the complexity of noise and the difficulty of extracting nodules. A deep learning model is introduced and a thyroid nodule detection method based on median filter and residual network is proposed. The median filtering method based on statistical threshold is used to improve the edge features of the nodules and realize the automatic enhancement of the ultrasonic image. Then, the CNN6-Residual model is constructed to extract and screen the characteristics of the nodules. It has cross layer connection and residual learning in order to reduce the difficulty of network training. The experimental results show that the accuracy of this method is 97.03%, which has high clinical application value.

Key words: statistical threshold, median filter, depth residual network, thyroid nodules, feature extraction

摘要： 利用超声图像对甲状腺结节进行检测在医学诊断中具有至关重要的作用。针对传统机器学习方法处理过程中存在噪声复杂、特征提取困难等问题，提出一种基于中值滤波和深度学习残差网络的甲状腺超声图像结节检测方法。采用统计阈值中值滤波方法，提高结节边缘特征，实现超声图像自动增强；构建CNN6-Residual模型提取和筛选结节特征，使用跨层连接和残差学习降低网络训练难度。实验结果表明，该方法检测准确率达到97.03%，具有较高的临床应用价值。

关键词: 统计阈值, 中值滤波, 残差神经网络, 甲状腺结节, 特征提取