Image Classification Of Osteoporosis Based On Transfer Learning Convolutional Neural Network

Objective:Based on the transfer learning method,the optimal model was obtained by comparing the performance of AlexNet,VGGNet and ResNet convolutional neural network models in the classification of osteoporosis images.Providing a more rapid and convenient diagnostic method for osteoporosis can improve the detection rate of bone mineral density,help the checkers to prevent bone loss and early development of osteoporosis,to promote its healthy development.Methods:In terms of data,424 images of osteoporosis patients from 2018 to 2021 were collected from two hospitals in Guiyang.This study mainly carried out the following steps for the classification task of osteoporosis:1.Data augmentation;2.Transfer learning convolutional neural network model construction;3.Parameter setting;4.Model comparison.In the data augmentation stage,random rotation,contrast adjustment,sharpness and the combination of these methods are used for augmentation.In the construction phase of the transfer learning convolutional neural network model,the convolutional neural network model with excellent performance in ImageNet Image classification contest is selected to build the model through the transfer learning method.In the parameter setting stage,the Learning Rate,Iteration Times and Batch-Size of AlexNet,VGGNet and ResNet were set equally,so as to prevent models from being unable to be properly compared due to different super parameters.Loss value and accuracy curve were made respectively.Record the accuracy and minimum loss of the model during training and testing for later comparison;Then the features are recorded by computer,the dimensions are reduced by Lasso,and the sigmoid activation function is used for prediction.In the model comparison stage,Accuracy,Precition,Receiver Operating Characteristic Curve(ROC),Area Under the Curve(AUC),F1-Score and time cost were selected to compare the three models.Results:AlexNet model:The accuracy rate(ACC)was 90.08%,the minimum LOSS(LOSS)was 0.2869,the subject area under the working curve(AUC)of the training set was 0.990,the subject area under the working curve(AUC)of the test set was 0.968,and the score of F1-Score was 0.9535.In terms of time cost,The total iteration time of 200 rounds was 7800 seconds,and the average time of each round was 39 seconds.ResNet model:The accuracy rate(ACC)was 98.78%,the minimum LOSS(LOSS)was 0.1587,the subject area under the working curve(AUC)of the training set was 0.998,the subject area under the working curve(AUC)of the test set was 0.975,and the score of F1-Score was 0.9714.In terms of time cost,The total iteration time of 200 rounds is 3600 seconds,with an average time of 18 seconds per round.VGGNet model:The accuracy rate(ACC)was 97.72%,the minimum LOSS(LOSS)was 0.1450,the subject’s area under the working curve(AUC)of the training set was 0.997,the subject’s area under the working curve(AUC)of the test set was 0.987,and the score of F1-Score was 0.9589.In terms of time cost,The 200 iterations took 5800 seconds,with an average time of 29 seconds per round.Conclusion:1.In terms of small-sample medical image data,it is feasible to enhance the original data by data enhancement method and classify the data set by transfer learning.2.AlexNet,VGGNet and ResNet all performed well in the task of osteoporosis classification under transfer learning.After the comparison of ACC,LOSS,ROC curve,AUC,F1-Score,time cost and other comprehensive evaluation indicators,ResNet performed best.