A Research On Prediction Method Of Mechanical Properties Of Trabecular Bone Based On Deep Learning

Osteoporosis is one of the common metabolic diseases.Biomechanical analysis of bone microstructure based on Micro-CT images is of great significance for the early prevention and diagnosis of osteoporosis.At present,bone mineral density measurement is used clinically as the gold standard for the diagnosis of osteoporosis,but studies have shown that bone mineral density can only explain 60% to 70% of bone strength changes,and trabecular bone microstructure is an important element that affects bone strength.In order to establish the connection between the trabecular bone morphology and bone strength,this paper has conducted an in-depth study on the prediction method of mechanical properties of trabecular bone based on deep learning.The main research contents are as follows:(1)A three-dimensional image data set of trabecular bone was produced.In this project,in cooperation with Gulou Hospital,9 femoral head samples were collected,and the samples were reconstructed by Micro-CT scanning;then,samples of the trabecular modulus prediction data set were obtained through image preprocessing steps such as region extraction,threshold segmentation and region cutting;after that,a cone-beam Micro-CT simulation system was built based on the ASTRA toolbox,and the high-resolution CT images were down-sampled by the simulation system to obtain the samples of the trabecular super-resolution modulus prediction data set.Based on the finite element analysis software Abaqus,the biomechanical analysis of the data set samples was carried out,the elastic modulus of the samples is calculated.After the data set is made,this paper cleans the data set according to the bone volume fraction of the sample,and then enhances the data set through rotation and translation transformations,increasing the diversity of the data set samples.(2)A variety of trabecular bone modulus regression networks were built,and the networks were trained and evaluated.First,this paper improves Vox Res Net and Dense Vox Net based on different types of residual modules,Dense modules,and SE modules,and builds 10 types of deep regression networks with different structures;then uses the training set obtained by dividing the trabecular modulus prediction data set to train the regression networks;after the training,the accuracy of the regression networks are evaluated through the test set.After comparing multiple indicators,it was found that SE-Dense Vox Net-B performed best on the test set.(3)Two prediction methods for trabecular bone super-resolution modulus are proposed,and the two methods are compared and evaluated.This article proposes two schemes,the first is to build a shallow regression network based on Vox Net,and the second is to use SE-Dense Vox Net-B as a pre-training model for transfer learning.Comparing the performance of the two schemes on the test set,we can see that the super-resolution modulus prediction method based on transfer learning is more accurate,and both schemes are superior to the calculation results of the finite element method.