首页 | 官方网站   微博 | 高级检索  
     

基于Transformer增强卷积的膝关节磁共振影像年龄预测
引用本文:朱昊哲,邓小冬,廖培希,杜文超,陈怀歆,刘洪,陈虎,邓振华,杨红雨.基于Transformer增强卷积的膝关节磁共振影像年龄预测[J].四川大学学报(自然科学版),2023,60(5):052001-101.
作者姓名:朱昊哲  邓小冬  廖培希  杜文超  陈怀歆  刘洪  陈虎  邓振华  杨红雨
作者单位:四川大学计算机学院;四川大学华西基础医学院与法医学院;成都市第六人民医院;四川大学视觉合成图形图像技术重点学科实验室;四川大学计算机学院; 四川大学视觉合成图形图像技术重点学科实验室
基金项目:四川省卫生健康委员会科研课题(19PJ007); 成都市卫生健康委员会科研课题(2022053);四川省自然科学基金(2022NSFSC1286); 成都市重点研发支撑计划项目(2021YF0501788SN)
摘    要: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.

关 键 词:深度学习  膝关节年龄预测  核磁共振影像  计算机辅助诊断
收稿时间:2022/11/30 0:00:00
修稿时间:2023/2/28 0:00:00

Transformer enhanced convolution based knee age estimation on MRIs
ZHU Hao-Zhe,DENG Xiao-Dong,LIAO Pei-Xi,DU Wen-Chao,CHEN Huai-Xin,LIU Hong,CHEN Hu,DENG Zhen-Hu,YANG Hong-Yu.Transformer enhanced convolution based knee age estimation on MRIs[J].Journal of Sichuan University (Natural Science Edition),2023,60(5):052001-101.
Authors:ZHU Hao-Zhe  DENG Xiao-Dong  LIAO Pei-Xi  DU Wen-Chao  CHEN Huai-Xin  LIU Hong  CHEN Hu  DENG Zhen-Hu  YANG Hong-Yu
Abstract: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.
Keywords:Deep learning  Knee age estimation  Magnetic resonance imaging  Computeraided diagnosis
点击此处可从《四川大学学报(自然科学版)》浏览原始摘要信息
点击此处可从《四川大学学报(自然科学版)》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司    京ICP备09084417号-23

京公网安备 11010802026262号