Design And Implementation Of Bone Age Prediction Based On Lightweight Convolutional Neural Network

Recently,the age of criminal responsibility in the juvenile law was lowered to 12,and the report of the Fourth Session of the 13 th National People’s Congress showed that in 2020,the number of prosecutors prosecuting minors suspected of crimes was as high as 33,000,which is a crime of a younger age in vicious cases.The trend has increased the domestic demand for juvenile bone age identification.In the domestic judicial field,bone age appraisal has played a huge role as evidence in the sentencing of suspects.At present,the domestic bone age identification mainly relies on manual identification,relying on the bone age scoring method to grade the hand bone development to get the predicted bone age.The disadvantage is that the relevant knowledge needs to be mastered in advance and cumbersome operations are required,which is relatively high threshold for non-professionals.The cost of time consumed is very high,which is not conducive to its popularization in life.In order to promote the popularization of bone age identification in real life,we must first solve four problems.The first is the labor cost of bone age identification,the second is the accuracy of bone age identification,the third is the equipment requirements for bone age identification,and the last is the speed of bone age identification.This thesis proposes an automatic bone age assessment scheme based on a lightweight neural network.The current mainstream YOLOv3-SPP detection framework is improved from the basic network,a priori frame generation,target frame loss,and multi-scale detection.The HYOLOv3 frame is proposed and designed.A high-efficiency and high-quality bone age assessment network Mullight Net specializes in the regression prediction of bone age.The method is to first perform traditional image preprocessing technology on the captured hand bone X-ray images to optimize and expand the quality of the data set,and then use the hand bone detection framework HYOLOv3 to extract the hand bone region of interest,and the extracted key regions are posture correction and combination.Later,the bone age is obtained through regression of the bone age assessment network Mul-light Net,which solves the above problems from the four aspects of automation,high precision,low computing power,and high speed.The main research contents of this thesis are as follows:(1)Research on the image preprocessing method of the RSNA hand bone public set,including the use of contrast enhancement and filtering noise reduction technology for the hand bone image to optimize the quality of the data set,and the removal of redundant information through traditional image segmentation algorithms.The above operations are used to expand the original data set(2)Improve the target detection framework YOLOv3-SPP and propose the HYOLOv3 framework.By studying the mainstream international and domestic bone age scoring methods,the region of interest that retains the characteristic information of the hand bones is selected to the greatest extent,and the existing public sets are trained and evaluated the weight of the model with excellent effect is designed as the input of the bone age assessment network after the posture correction of the ROI area combination three network scheme.(3)Design and train the bone age parallel evaluation network Mul-light Net,designed comparative experiments from multiple angles to prove the superiority of the evaluation scheme in this thesis,and compared the final result with multiple similar bone age evaluation studies at domestic and abroad to prove its possession Higher accuracy,lower memory footprint,and faster evaluation speed.(4)Design and develop an automatic bone age assessment system that integrates the detection and evaluation algorithm interface of this article,which provides a simple,fast,highprecision,low-time-consuming bone age prediction,and the user history can be recalled through the system,which verifies that the research is the feasibility of promotion in the real world.This thesis also developed an event-driven high-concurrency server framework during the system implementation process,which performed well in the actual test process.