Deep Segmentation And Radiomics Feature Extraction Of Bone Metastases

Single Photon Emission Computed Tomography(SPECT)realizes the imaging of the whole body bone morphology and metabolism with the help of single photon nuclide markers,and realizes the noninvasive detection of the disease by finding abnormal radioactive concentration or sparse defect in the lesion site.Currently,it has become the preferred method to diagnose bone metastasis of malignant tumors.Due to the limitation of imaging technology and imaging equipment,SPECT bone scan images often have shortcomings such as low resolution and blurred boundary of focus area,which makes the diagnosis work time-consuming,low efficiency and accompanied by subjective errors.Furthermore,accurate segmentation of focal areas in bone scan images is particularly important.With the wide application of deep learning in the field of computer vision,in particular,convolutional neural network can extract image context information in the form of multi-layer receptive field without human participation,which opens up a new way for automatic segmentation of medical images.Based on deep learning method and Radiomics method,this paper studied the depth segmentation of bone metastases and their Radiomics feature extraction,and mainly carried out the following three aspects of research work.(1)Deep segmentation of bone metastases.Deep segmentation is the deep learning method to achieve focal region segmentation.In this paper,an end-to-end segmentation network composed of feature extraction and pixel classification is constructed.In the stage of feature extraction,the depth supervision idea of multi-scale fusion is proposed.In the feature classification stage,traditional Support Vector Machines(SVM)were used to classify the extracted features.The results showed that DSC,Recall and CPA were 0.6556,0.6257 and 0.6885,respectively.Then,a dual-path network consisting of image segmentation network and image reconstruction network is proposed to present the segmentation results to the reconstructed images,so as to assist doctors to make clear clinical decisions.In the segmentation network,the Attention R-UNet structure was proposed,and the composite loss function was constructed to eliminate the imbalance of positive and negative samples and improve the classification of difficult samples.The results showed that DSC,Recall and CPA were 0.6610,0.6381 and 0.6857,respectively.Compared with the classical model,the results of the proposed model are optimal.(2)Radiomics feature extraction and classification of bone metastases.The segmented lesion area was taken as the research object.By extracting a large number of radiomics features from the lesion area to describe information such as tumor biological characteristics and heterogeneity,the classification of adenocarcinoma and small cell carcinoma is realized.First,the lesion area was acquired from the whole body SPECT image to construct a radiomics datasets.Secondly,the radiomics algorithm is used to extract 93-dimensional radiomics features containing intensity and texture,and different feature filtering methods are used to delete irrelevant and redundant feature parameters to retain the strongly correlated features of the classification target.Finally,a support vector machines(SVM)parameter selection method based on Genetic Algorithm(GA)is proposed,and a GA-SVM classifier is established to classify the screened features.The experimental results show that Acc,Prec,Rec,and F-1 are 0.7338,0.7339,0.7339,and 0.7339,respectively,which are the best results compared with other classifiers.