基于深度学习区域融合的骨龄评价研究

Research on Bone Age Evaluation Based on Fusion of Key Regions in Deep Learning

骨龄判断在医学、司法鉴定、体育等领域得到广泛应用,人工评估费时且结果可能因医师水平不同出现差异。提出一种基于深度学习卷积神经网络和手骨数据关键区域融合、拓展的方法对手骨X线片骨龄自动评价,快速得到手骨X线片的准确骨龄。方法收集四川大学华西第二医院及RSNA2017骨龄比赛X线片数据,预处理后,设计深度学习骨龄评价神经网络进行训练,获取模型对骨龄的预测值与真实值误差在±1岁误差的准确率和平均绝对误差(MAE),结果显示四川大学华西第二医院数据误差±1岁的准确率:女性为94.25%,男性为94%;平均绝对误差:女性为0.5125岁,男性为0.5575岁。RSNA数据误差±1岁的准确率:女性为93%,男性为95%;平均绝对误差:女性为0.56岁,男性为0.55岁。通过广泛的对比,研究提出的方案不仅得到更高的骨龄评价准确率,同时研究提出的卷积神经网络结构简单,在网络训练和评价速度方面均具有明显优势。

Bone age judgment has been widely used in the fields of medicine,forensic identification,sports,etc.,but the manual evaluation is timeconsuming and the results may differ due to different levels of physicians.This paper proposes a method based on deep learning convolu tional neural network and key area fusion and expansion to automatically evaluate the bone age of the X-ray film of the hand bone which can get accurate bone age of hand bone X-ray film in little time.Methods Collect X-ray data of Sichuan University West China Second Hospital and RSNA2017 bone age competition,then design a deep learning bone age evaluation neural network to training these data after pre-processing them,and get the ratio of the predicted bone age and true bone age errors in 1 year old and The mean absolute error(MAE).The results showed that the ratio of the data error in 1 year old of the Second Hospital of West China University of Sichuan Univer sity was 94.25%for female and 94%for male.The average absolute error was 0.5125 years old for female and 0.5575 years old for male.The ratio of the RSNA test set error in 1 year old:93%for female and 95%for male;Mean absolute error:0.56 for female and 0.55 for male.Through extensive comparison,the proposed scheme not only obtained higher accuracy of bone age evaluation,but also the proposed convo lutional neural network has a simple structure and obvious advantages in network training and evaluation speed.