Study On Deep Learning Based Detection And Representation Of Bone Imaging Lesions

SPECT whole body bone imaging is the most common diagnostic method in the field of nuclear medicine.It has the potential to capture the functional status and structural morphology of organs,tissues and other body parts at the same time.It plays an irreplaceable role in the accurate assessment and staging of diseases.Reliable segmentation of the lesion area in SPECT bone images is helpful to calculate the diagnostic indicators of drug uptake and metabolic distribution in tumor location.The rapid development of deep learning technology,especially the hierarchical feature extraction function of convolutional neural network,has laid a foundation for the realization of semantic segmentation of objects in images.Taking low resolution SPECT bone scan image as the research object,this thesis studies the automatic segmentation of lesions based on the combination of traditional machine learning and deep learning.The main work and results are as follows:(1)Three traditional machine learning algorithms and FCN model based on deep learning are proposed to segment bone metastases in SPECT bone imaging.Firstly,the chest region is extracted from the whole body SPECT bone images,and the imaging data are normalized,and the images are mirrored,translated and rotated to expand the data set;Then,three traditional machine learning algorithms based on K-means clustering algorithm,region growing algorithm and C-V model,and the segmentation method of FCN model based on deep learning are constructed;Finally,the focus segmentation was performed on a real SPECT bone images,the Tanimoto similarity coefficients of the three machine learning algorithms were 0.7303,0.7768 and 0.8076,respectively,and the CPA,MPA and Io U of FCN model were 0.654,0.609 and 0.453,respectively.The experimental evaluation shows that the traditional machine learning algorithm is more effective than FCN model for SPECT bone metastases,which can provide more auxiliary information for doctors’ diagnosis.(2)The segmentation models based on deep learning are proposed to segment multiple lesions in SPECT bone imaging.Firstly,the chest and joint region are extracted from the whole body SPECT bone scan images,and the imaging data are normalized and expanded;Then,the segmentation models based on Mask R-CNN,U-Net and the improved automatic segmentation SPECT bone imaging lesion model called R＿U-Net based on U-Net are studied and constructed.In this model,the residual block is introduced to replace the common convolution layer to deepen the depth of U-Net network,Dice loss coefficient is introduced as the loss function to overcome the class imbalance problem,so as to improve the segmentation accuracy of the model;Next,the focus segmentation was performed on a real SPECT bone images,the values of CPA and IOU in bone metastases of R＿U-Net model were 0.772 and0.610,respectively,and the values of CPA and IOU in arthritis lesions were 0.688 and0.648,respectively;Finally,the location,shape,level,state and category of the segmented lesions were extracted.The experimental results show that the segmentation accuracy of bone metastases and arthritis lesions is improved in R＿U-Net model compared with U-Net model.(3)Based on the three traditional machine learning methods and four deep learning methods proposed in this thesis,a lesion segmentation system for SPECT bone imaging is designed and implemented.The system has the basic functions of data upload,display and save,and provides a variety of preprocessing methods such as normalization and geometric change,and medical image segmentation evaluation index,At the same time,by selecting different segmentation methods,the corresponding SPECT image segmentation results can be displayed,which can basically meet the needs of users.