Ultrasound image-based thyroid nodule automatic segmentation using convolutional neural networks
Ultrasound image-based thyroid nodule automatic segmentation using convolutional neural networks
复制标题
使用卷积神经网络基于超声图像的甲状腺结节自动分割
DOI：
10.1007/s11548-017-1649-7
复制
发表时间：
2017-11-01
影响因子：
3
通讯作者：
Kong, Dexing
中科院分区：
文献类型：
--
作者：
Ma, Jinlian;Wu, Fa;Kong, Dexing
目的从超声图像中提取甲状腺结节边界，对临床指标的计算和甲状腺疾病的诊断具有重要意义。然而，由于甲状腺结节的形态和成分与背景相似，对其进行准确的自动分割是一项具有挑战性的工作。本文采用深度卷积神经网络(CNN)对超声图像中的甲状腺结节进行自动分割。方法将甲状腺结节分割问题描述为一个忽略块间关系的块分类问题。具体来说，CNN使用来自正常甲状腺和甲状腺结节图像的图像块作为输入，然后生成分割概率图作为输出。多视点策略被用来提高基于CNN的模型的性能。实验结果表明，本文提出的方法在甲状腺结节分割上优于已有的方法。实验结果表明，基于CNN的模型能够准确有效地提取出甲状腺超声图像中的多个结节。基于CNN的模型可以获得总体折叠的重叠度、骰子比、真正确率、假正确率和修正的Hausdorff距离的平均值，分别为、。定量结果也表明，我们的方法是如此高效和准确，足以很好地取代耗时和繁琐的人工分割方法，展示了潜在的临床应用。
PurposeDelineation of thyroid nodule boundaries from ultrasound images plays an important role in calculation of clinical indices and diagnosis of thyroid diseases. However, it is challenging for accurate and automatic segmentation of thyroid nodules because of their heterogeneous appearance and components similar to the background. In this study, we employ a deep convolutional neural network (CNN) to automatically segment thyroid nodules from ultrasound images.MethodsOur CNN-based method formulates a thyroid nodule segmentation problem as a patch classification task, where the relationship among patches is ignored. Specifically, the CNN used image patches from images of normal thyroids and thyroid nodules as inputs and then generated the segmentation probability maps as outputs. A multi-view strategy is used to improve the performance of the CNN-based model. Additionally, we compared the performance of our approach with that of the commonly used segmentation methods on the same dataset.ResultsThe experimental results suggest that our proposed method outperforms prior methods on thyroid nodule segmentation. Moreover, the results show that the CNN-based model is able to delineate multiple nodules in thyroid ultrasound images accurately and effectively. In detail, our CNN-based model can achieve an average of the overlap metric, dice ratio, true positive rate, false positive rate, and modified Hausdorff distance as,,,,on overall folds, respectively.ConclusionOur proposed method is fully automatic without any user interaction. Quantitative results also indicate that our method is so efficient and accurate that it can be good enough to replace the time-consuming and tedious manual segmentation approach, demonstrating the potential clinical applications.