Objective

Aiming at the time consuming processing of existing thyroid nodule detection and difficulty in feature extraction, U-Net-based thyroid nodule detection is proposed to perform computed aided diagnosis.

Method

This paper proposes a mark-guided ultrasound deep network segmentation model of thyroid nodules. By comparing with VGG19, Inception V3, DenseNet 161, segmentation accuracy, segmentation edge and network operation time, it is found that the algorithm in this paper has relative advantages.

Results

U-Net network-based ultrasound thyroid nodules segmented the nodule area overlapped with the manually depicted nodule area close to 100%, the segmentation accuracy rate was as high as 0.9785, and the U-Net segmentation result was closer to the manually depicted nodule. The accuracy of U-Net segmentation of the thyroid is about 3% higher than the other three networks.

Conclusion

The segmentation of nodules based on U-Net proposed in this paper significantly improves the segmentation accuracy of thyroid nodules with a small training data set, and provides a comprehensive reference for clinical diagnosis and treatment.

中文翻译：

---

基于U-Net网络的超声甲状腺结节检测方法

目的

针对现有甲状腺结节检测的费时处理和特征提取困难的问题，提出了基于U-Net的甲状腺结节检测进行计算机辅助诊断的方法。

方法

本文提出了一种基于标记的甲状腺结节超声深层网络分割模型。通过与VGG19，Inception V3，DenseNet 161，分割精度，分割边缘和网络运行时间进行比较，发现本文算法具有相对优势。

结果

基于U-Net网络的超声甲状腺结节对与手工绘制的结节区域重叠的结节区域进行了分割，接近100％，分割的准确率高达0.9785，并且U-Net的分割结果更接近于手工绘制的结节。甲状腺的U-Net分割准确性比其他三个网络高约3％。

结论

本文提出的基于U-Net的结节分割方法，只需少量训练数据集，就可以显着提高甲状腺结节的分割精度，为临床诊治提供全面的参考。