Automated segmentation of the knee for age assessment in 3D MR images using convolutional neural networks
Automated segmentation of the knee for age assessment in 3D MR images using convolutional neural networks
复制标题
DOI：
10.1007/s00414-018-1953-y
复制
发表时间：
2019-07-01
影响因子：
2.1
通讯作者：
der Mauer, Markus Auf
中科院分区：
文献类型：
--
作者：
Proeve, Paul-Louis;Jopp-van Well, Eilin;der Mauer, Markus Auf
年龄评估用于估计缺乏法律的文件的个人的实际年龄。近年来的研究表明，膝关节生长板的骨化程度与青少年和青壮年的实际年龄有关。为了验证这一假设，需要分析大量的数据集。一种能够自动检测和分析骨骼结构的方法对于处理大型数据集可能是必要的。本研究的目的是使用卷积神经网络开发基于3D MR图像的全自动2D膝关节分割。共有76个数据集可用，并分为训练集（74%），验证集（13%）和测试集（13%）。应用多个预处理步骤来校正图像强度值并减小图像尺寸。采用图像增强来虚拟地增加用于训练的数据集大小。分割任务的建议架构类似于用于U-Net的编码器-解码器模型类型。与手动分割相比，经过训练的网络实现了98%的骰子相似性系数得分和96%的交集。该模型的查准率和查全率达到了平衡，误差仅为1.2%。在训练过程中没有观察到过拟合。作为概念证明，预测的分割用于145名受试者的年龄估计。初步结果表明，这种方法的潜力达到0.48 +/-0.32年的测试集的14个科目的平均绝对误差。所提出的自动分割可以有助于更快，可重复和潜在的更可靠的年龄估计在未来。
Age assessment is used to estimate the chronological age of an individual who lacks legal documentation. Recent studies indicate that the ossification degree of the growth plates in the knee joint correlates with chronological age of adolescents and young adults. To verify this hypothesis, a high number of datasets need to be analysed. An approach which enables an automated detection and analysis of the bone structures may be necessary to handle large datasets. The purpose of this study was to develop a fully automatic 2D knee segmentation based on 3D MR images using convolutional neural networks. A total of 76 datasets were available and divided into a training set (74%), a validation set (13%) and a test set (13%). Multiple preprocessing steps were applied to correct image intensity values and to reduce the image size. Image augmentation was employed to virtually increase the dataset size for training. The proposed architecture for the segmentation task resembles the encoder-decoder model type used for the U-Net. The trained network achieved a dice similarity coefficient score of 98% compared to the manual segmentations and an intersection over union of 96%. The precision and recall of the model were balanced, and the error was only 1.2%. No overfitting was observed during training. As a proof of concept, the predicted segmentations were used for the age estimation of 145 subjects. Initial results show the potential of this approach attaining a mean absolute error of 0.48 +/- 0.32 years for a test set of 14 subjects. The proposed automated segmentation can contribute to faster, reproducible and potentially more reliable age estimation in the future.