基于Transformer增强卷积的膝关节磁共振影像年龄预测

Transformer enhanced convolution based knee age estimation on MRIs

年龄预测是临床医学中的一个重要课题和非常活跃的研究领域.最近,由于传统影像学检查中电离辐射的缺点,越来越多的研究使用磁共振影像进行年龄预测.本文基于膝关节MRI数据集,提出了一种新的端到端网络,结合卷积神经网络和Masked-Transformer网络互补地来提取局部特征和全局依赖,并使用一个特征聚合模块来聚合不同局部膝关节MRI切片的特征.通过整合卷积神经网络的特征图和视觉Transformer分支的特征编码,特征提取模块可以互补地提取局部和全局信息,更好地提取与年龄相关的特征.同时,该网络使用由图注意力网络组成的特征聚合模块,用于在特征级别集成不同MRI切片的局部特征,实现多切片局部特征之间的交互.大量实验表明,该方法可以在膝关节MRI年龄估计任务中达到最先进的性能.具体而言,本文方法在MRI数据集上进行了测试,该测试集包括44个年龄在12.0~25.9岁之间的膝关节MRI样本,其中五折交叉验证的最佳结果是年龄平均绝对误差为1.57±1.34岁.

Age estimation is regarded as a crucial topic and a very active research field in clinical medicine.Recently,due to the drawback of ionizing radiation from the traditional imageological examination,growing more and more studies have focused on using magnetic resonance imaging(MRI)for bone age prediction.This paper proposes a novel end-to-end network based on the knee MRI dataset,which combines the convolution neural network(CNN)and Masked-Transformer network to extract complementary features,and uses a feature aggregation module to aggregate features of different local knee MRI slices.By integrating the feature maps of CNN and the patch embeddings of visual transformer branches,the feature extraction module can complementarily acquire local and global information to better extract age-related features.A feature aggregation module composed of the graph attention network is proposed in our work to integrate the local features of different MRI slices at the feature level to achieve the interaction between multiple slice features.Extensive experiments demonstrate that our method can achieve state-of-the-art performance in the knee MRI age estimation task.Specifically,our method is tested on a dataset including 44 knee MRI samples aging from 12.0 to 25.9 years,and the best result of five-fold cross-validation is a mean absolute error of 1.57±1.34 years in age regression.