甲状腺结节超声图像的分割方案

Segmentation Scheme for Ultrasound Images of Thyroid Nodules

摘要: 目的：通过计算机辅助技术(CAD)来进行甲状腺超声图像处理，显示甲状腺结节良恶性的细微差别，来为缓解临床上甲状腺医师的鉴别诊断压力。方法：选取流行的开放数据集DDTI (Digital Database of Thyroid Image)进行预处理，在级联框架中训练stage 1和stage 2两个超参数相同的网络，之后进行五折交叉验证进一步验证CNN模型，然后在未经处理的原始图像上测试图像。最后合并得到集合预测结果。结果：有效地避免在过拟合的情况下实现了分割，交并比为80.91%。结论：GT (Ground Truth，真实度)各自训练两个网络。然后inference的时候用stage 1框出小区域，stage 2在这个小区域上希望得到更细致的分割结果，这种方法确实对甲状腺图像的分割结果有效果。

Abstract: Objective: To perform thyroid ultrasound image processing by computer-aided technology (CAD) to show the subtle differences between benign and malignant thyroid nodules to ease the differential diagnosis of thyroid physicians in clinical practice. Two networks with the same hyperparameters of stage 1 and stage 2 were trained in a cascade framework, followed by a five-fold cross- validation to further validate the CNN model, and then the images were tested on the unprocessed original im-ages. Finally, the combined prediction results were obtained. Results: The segmentation was effec-tively achieved while avoiding overfitting, and the cross-comparison ratio was 80.91%. Conclusion: GT (ground truth) was trained for each of the two networks. Then inference was done with stage1 to frame a small region, and stage 2 was used on this small region hoping to get more detailed seg-mentation results, this method did improve the segmentation results of thyroid images.