基于深度学习的智能骨龄分类器  被引量:3

Intelligent Bone Age Classifier Based on Deep Learning Algorithm

在线阅读下载全文

作  者:郭子昇 王吉芳[1] 苏鹏 GUO Zi-Sheng;WANG Ji-Fang;SU Peng(Mechanical Electrical Engineering School,Beijing Information Science and Technology University,Beijing 100192,China)

机构地区:[1]北京信息科技大学机电工程学院,北京100192

出  处:《计算机系统应用》2022年第6期339-346,共8页Computer Systems & Applications

基  金:国家自然科学基金(52005045);北京市自然科学基金-海淀原始创新联合基金(L192018);北京高校高精尖学科建设项目(77D2111002)。

摘  要:通用深度学习算法提取的医学手骨图像特征不能很好地区分相近年龄图像的差异,这导致骨龄分类器的预测精度较低.根据基于深度学习的轻量级神经网络MobileNet设计了一种改进的骨龄分类器RIL-MobileNetV3 Large,通过改进LBP处理层得到了具有细致纹理特征的手骨数据集并引入注意力机制进行自动定位,通过学习处理层处理后的手骨X光片中的深层区域特征完成识别和骨龄的分类,在公共数据集上进行实验并对该分类器进行多次训练调优,结果表明改进设计的分类器在骨龄预测中具有高达94.204%的准确率和0.350岁的均值误差,而且改进的轻量级网络为可移动智能便携预测骨龄奠定基础.The extracted features of medical hand bone images by the general deep learning algorithm can’t well distinguish the differences from images with similar age. It leads to the low prediction accuracy of bone age classifier. An improved bone age classifier, named RIL-MobileNetV3 Large, in accordance with the deep learning-based lightweight neural network MobileNet is designed. A dataset of hand bone is generated by the improved LBP processing layer with fine textures and an attention mechanism for automatic positioning is introduced. It complete the recognition and classification of bone age by learning deep area features in the X-ray of hand bone treated by the processing layer. A lot of training is carried out for tuning accompanied by the experiment on public datasets. The results show that the improved classifier has got a high accuracy of 94.204% and a mean error of 0.350 years in the bone age prediction. The improved lightweight network lays a foundation for mobile, intelligent and portable prediction devices of bone age.

关 键 词:骨龄分类器 深度学习 LBP纹理增强 注意力机制 MobileNet 神经网络 纹理特征 

分 类 号:R318[医药卫生—生物医学工程] TP18[医药卫生—基础医学] TP391.41[自动化与计算机技术—控制理论与控制工程]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象