A Preliminary Study Of Deep Learning Model In The Diagnosis Of Osteoporosis

Objective:To explore the application of deep learning in the diagnosis of osteoporosis,and to establish a deep learning model for osteoporosis classification based on quantitative computed tomography(QCT)combined with laboratory parameters.To explore whether the combination of laboratory parameters can improve the classification performance of osteoporosis in a deep learning model compared to the use of images alone.Method:A retrospective study was conducted on 1450 patients who underwent QCT examination in Dazhou Central Hospital from January 2016 to December 2020.Among them,611 patients were used as the training set and 153 patients as the validation set.Our proposed approach is based on two parts.Firstly,the QCT image is automatically segmented by full convolutional neural network U-NET.Then,using the T score of quantitative computed tomography(QCT)as the reference standard,the convolutional neural network(CNN)DenseNet-121 was used to classify osteoporosis according to the QCT images.Support vector machine(SVM)and fuzzy neural network(FNN)were used to classify osteoporosis according to laboratory parameters.Finally,a deep learning model containing images and laboratory parameters was constructed.The performance of the deep learning model was evaluated through accuracy,recall rate(sensitivity),accuracy and F1 score.Results:In the end,764 patients were enrolled in the study,according to the proportion of 4:1 divided the patients into training set(N=611)and test set(N=153).For the classification of osteoporosis,the accuracy(0.9021),recall rate(0.9358)and F1 score(0.9358)of the two classes experiment(osteoporosis and bone loss)combined with laboratory parameters were the highest,and the model obtained the best performance.Conclusion:Studies have shown that QCT images can be used directly to classify osteoporosis based on deep learning model,and the classification performance of the model can be further improved by adding laboratory parameters.