Research On Deep Learning-Based Opportunistic Screening Model For Osteoporosis In 3D CT Images

Osteoporosis(OP)is a highly prevalent and potentially debilitating disease.While dual-energy X-ray absorptiometry(DXA)is considered the "gold standard" for OP diagnosis by measuring bone density,the lack of public awareness and limited preventive and diagnostic capabilities at the primary healthcare level have resulted in low DXA testing rates.This has severely affected the diagnosis and treatment of osteoporosis patients.To address this issue,this thesis proposes a CT image-based opportunistic screening model for OP.By conducting OP screening on a large number of abdominal CT images of patients with other diseases,without additional radiation exposure,time,or cost,this model is of great significance for OP prevention and treatment.The main contributions of this thesis are as follows:(1)To address the lack of publicly available CT image-based OP screening datasets,we collected medical data and created a dataset of 618 cases containing CT images,vertebral body segmentation annotations,and DXA bone density annotations,which will be used for subsequent model training.(2)In order to improve the segmentation accuracy of L1-L4 lumbar vertebral body segmentation and enhance the classification performance of the screening model,this thesis proposed a two-stage segmentation model to address the problem of low accuracy of current OP screening models based on 2D U-Net network segmentation.The proposed model first performs coarse segmentation of lumbar vertebral body CT images and then refines the segmentation of the lumbar vertebral body edge.The model is based on the 3D U-Net and introduces residual and attention modules,and improves the loss function to improve the segmentation performance of the network.The experimental results show that the proposed model performs well in the L1-L4 lumbar vertebral body segmentation task,with a Dice coefficient of 0.9728 and an Average Symmetric Surface Distance(ASD)reduced to 0.10mm,which can obtain accurate three-dimensional segmentation results of lumbar vertebral bodies.(3)In order to improve the robustness and accuracy of current OP screening models,this thesis proposed a deep learning-based OP screening model,which improves the classification performance of the model by studying the region of interest and improving the classification network.The model extracts complete three-dimensional CT images of L1-L4 lumbar vertebral bodies as the region of interest based on the segmentation results of lumbar vertebral bodies.Compared with only using trabecular bone CT images,the Area Under Curve of ROC(AUC)of the Receiver Operating Characteristic(ROC)curve is increased by 0.046.In addition,the model introduces attention modules in the 3D DenseNet network to further improve the classification performance.Finally,to solve the problem of overfitting in the model training process,transfer learning is used for pre-training to make the model’s loss function more stable.The experimental results show that the proposed OP screening model has an AUC of 0.929,which can effectively classify normal population,patients with decreased bone mass,and patients with osteoporosis.In summary,the deep learning-based three-dimensional CT image opportunistic OP screening model proposed in this thesis improves the problems of low segmentation accuracy,single classification features,and improper region of interest extraction in current screening models through improvements in the segmentation and classification models,and enhances the classification performance of the screening model.