Aided Detection Of Bone Age In Hand X-ray Films Based On Deep Learning

BA is an abbreviated form of bone age,it is able to response human's growing condition and the extent of mature.The wrist can be a major part of bone age assessment,because the growing condition of wrist bone can represent the bone condition of whole body and human suffer the least radiation while taking X-ray film of wrists.Bone age assessment is widely used in clinic medicine,forensic medicine,sports medicine and so on.It is always a popular topic for international and domestic researchers by using X-ray film of hand bone for bone age assessment.Due to the disadvantages of traditional bone age evaluations,this article presented the method of automated bone age assessment based on deep learning,Inception V3 and Convolutional Block Attention Module(CBAM).The article explored the image pretreatment of hand bone X-ray films,the training of neural network model,the visualization of bone age assessment results and the design of remote bone age aided prediction platform.The major work and innovation points was listed as followings:(1)This article used the new pretreatment method of hand bone X-ray films,which used image segmentation and then image enhancement.Hand bone X-ray film can become high quality data sets by using this new pretreatment method,which can produce the superior bone age aided prediction model.Original image of data sets can make some problems,such as improper exposure,big differences of grayscale of images,the noise of labels and artifacts on the background.So it is not able to give the pleasant result by whole image enhancement.This study used u-net for image segmentation to remove the background of hand bone images,and then used the partial gray-level histogram for the treatment of part of hand,and finally produced images of the same size,background elimination,and contrast balanced of hand bone.(2)This study created two innovated points based on Inception V3 neural network model to achieve high-accuracy aided bone age prediction: 1.In consider of the significant differences of bone growing condition of different gender,gender was input to network based on image pixel.And regression prediction structure was instead of classification prediction structure.2.A new network structure based on Inception-CBAM was designed with the combination of CBAM and Inception V3 to enhance the prediction ability of network.Final results showed that the new structure of network in this study had a better performance on test sets.(3)In order to understand the important area of hand bone directly by revised network prediction model.This study took example by Grad-cam method which was used for classification.The modified method was used in regression test,which produced visual thermodynamic chart explainable neural network.The important area of hand bone can be showed to inexperienced doctors by this thermodynamic chart,which is able to help doctors predict results more quickly and accurately.(4)Bone age prediction model was designed and arranged to the cloud,so that doctors from remote regions can process can process bone age judgment by network conveniently.The simple application development based on Web layer was tested by Flask,which provided users a simple website to upload images,get diagnosis results and produce regional thermodynamic chart.