首页|基于细粒度特征提取的轻量骨龄评估级联方法

基于细粒度特征提取的轻量骨龄评估级联方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对现有骨龄自动评估模型细节特征提取能力较弱且没有兼顾手骨发育特点的问题，提出了一种基于轻量级网络的两阶段骨龄评估级联模型。首先，定位提取手掌区域进行背景去噪。其次，提出了一种新型的轴向空间注意力与多尺度并行空洞卷积相结合，融入模型中，从而提升模型细节特征提取能力。同时，将每个骨龄标签转化为两点分布向量，充分利用个体手骨发育的信息。实验结果表明，模型预测结果的平均绝对误差为 4。80 且评估误差在±0。5、±1、±2 岁以内的准确率分别为 69。26%、95。60%、99。80%。本研究提出的模型不仅能快速准确的评估骨龄，且充分考虑了手骨发育过程中所包含的信息，同时，基于轻量化的架构为其后续推广应用提供了基础，具有良好的临床应用前景。
A Cascade Method for Lightweight Bone Age Assessment Based on Fine-Grained Feature Extraction
A two-stage bone age assessment cascade model is proposed based on a lightweight network to address the issue of weak feature extraction ability and lack of consideration for hand bone development characteristics in ex-isting automatic bone age assessment models.Firstly,the hand area was positioned and extracted for background de-noising.Secondly,axial spatial attention proposed in this paper with multi-scale parallel null convolution was added to the model to enhance the detail feature extraction ability.Meanwhile,each bone age label was transformed into a two-point distribution vector,which considers the characteristics of individual skeletal development.The experimental results show that the mean absolute error of the model prediction results is 4.80 and the accuracy of the assessment error within±0.5、±1、±2 years is 69.26%,95.60%,and 99.80%,respectively.The model proposed in this study not only can quickly and accurately assess bone age and fully consider the information contained in the process of hand bone development,but also provide a basis for its subsequent application on account of its lightweight architec-ture,which has promising prospects for clinical application.

Axial spatial attentionMulti-scale dilated convolutionTwo-point distribution of bone age labelsCascade modeLightweight

李南欣、张俊然、程勃超

展开 >

四川大学电气工程学院,四川 成都 610065

四川大学华西第二医院,四川 成都 610041

轴向空间注意力 多尺度空洞卷积 两点分布骨龄标签 级联模型 轻量化

四川省科技厅重点研发项目智能电网四川省重点实验室应急重点项目成都市科技计划德阳科技局(揭榜挂帅)项目

2022YFS0178020IEPG-KL-20YJ012021-YF05-D0916-SN2021JBJZ007

2024

计算机仿真
中国航天科工集团公司第十七研究所

计算机仿真

CSTPCD
影响因子：0.518
ISSN：1006-9348
年,卷(期)：2024.41(4)
  • 21