Bone Age Assessment Of Children Based On Deep Learning

Bone age,as the biological age of humans,is an accurate reflection of an individual’s level of growth and physical condition.The bone age assessment can assist doctors for diagnosing children’s growth problems.Owing to the over-reliance on the experience of the doctor,the traditional bone age assessment from left-handed X-ray images may be subjective.With the development of computer vision in medical imaging,the deep learning application of bone age assessment has important research implications for children.However,existing bone age assessment methods based on deep learning only use object detection models to output bone age,ignoring problems such as detailed difference information and imbalanced sample distribution of bone age.In this thesis,based on the analysis of the bone age datasets,the bone age assessment is investigated by using deep learning methods.The contributions are as follows:1.A two-stage ordered bone age assessment method from coarse to fine is proposed to improve the accuracy of bone age assessment by using the detailed difference information of bone age images.Firstly,a children’s bone age dataset is collected,which contained 2518 left hand X-ray images that labelled by professional doctors.Then,aiming to classify bone classes and grades,a two-stage method of bone age assessment is proposed.Specifically,the first stage of the method uses the YOLOv5 model to detect the 13 bones in the bone images and recognize the bone classes.Then based on context-aware attention pooling,the second stage proposes a fine-grained classification model to further classify the bone grades.In addition,ordered regularization was used to mine the ordinality of the bone grade labels to further optimize the bone grade classification performance.Finally,on both clinical dataset and the North American Radiological Society(RSNA)dataset,the results show that the proposed method exploits the detailed difference information of bone age images to increase the accuracy of bone age assessment,achieving a mean absolute value error of6.28 months on the RSNA dataset.2.To relieve imbalanced distribution of bone age dataset,a bilateral-branch bone age classification method is proposed that fuses the plug-in module and class-balanced loss.Firstly,a bilateral-branch model is designed with fused plug-in module,which uses conventional and rebalanced learning bilateral-branch for bone age image classification.The method enhances feature extraction in strongly discriminated regions,utilizing weakly supervised selectors and combiners in each convolutional layer of the feature extraction module.Then the model uses graph convolution to fuse the enhanced features for improving the classification performance of imbalanced bone age images.Secondly,to raise the classification accuracy of imbalanced bone age images,a reweighted class-balanced loss is applied to upgrade the classification performance of classes with few numbers of samples.Experimental results on both clinical and RSNA datasets show that the proposed method performs well on bone age classification,especially on classes with few numbers of samples.3.In order to apply the work of this thesis in the physician’s diagnosis,a bone age assessment system for children is designed and implemented.In the bone age assessment system,the algorithm proposed in this thesis was applied and the user interface was designed using the Py Qt5 framework.This system uploads left hand Xray images to output 13 bone grades and bone age assessment results.The thesis has 37 figures,18 tables and 82 references.