Study On Automatic Diagnosis Of Bone Metastases By Classifing Bone Scan Images

With the development of deep learning technology and the in-depth application of medical imaging technology,the use of deep learning to analyze and process medical images has become a research hotspot in academia.Nuclear medicine image analysis based on computer-aided diagnosis technology can not only achieve automatic detection of diseases,but also it has the potential to improve the diagnostic efficiency and accuracy.However,unlike natural images and structural medical images,SPECT bone imaging is a typical ultra-low-resolution large-scale imaging modality.In addition,large-scale SPECT data collection is difficult and image labeling is time-consuming and labor-intensive.Automatic classification of whole-body bone scan images poses a great challenge.In order to build a reliable deep learning SPECT bone scan image classification model,in view of the above problems,this thesis studies the automatic image classification technology for automatic diagnosis of bone metastases,which involves regional segmentation of SPECT whole body bone imaging images,and disease detection-oriented methods.The construction of image binary classification model and the construction of image multi classification model for disease diagnosis.In general,this paper mainly carries out the following research work:(1)Regional segmentation of SPECT whole-body bone scan images.By applying statistical analysis method and combining the knowledge of human body morphology,a segmentation method of human body parts for SPECT whole body bone imaging is studied and proposed.First,the non-background “pixel” statistics are performed on the whole-body bone scan data;then,the segmentation feature pixel points are determined by applying polynomial curve fitting technology combined with the symmetrical features of the human body structure;finally,the shoulder segmentation points are empirically obtained and analyzed.Extract the thoracic bone region.The proposed segmentation method is experimentally verified using real clinical data,and the results show the feasibility and effectiveness of the proposed method.(2)Image binary classification for disease detection.By constructing deep classification models with different structures,the two-class disease detection of non-fused images and fused images for bone metastases is studied and proposed.First,extract the region of interest from the whole-body bone scintigraphy image by searching for feature segmentation points,and perform personalized data enhancement;then,build a non-fused classification model and a fusion classification model based on the sequence structure and non-sequence structure,and at the same time The parameters are fine-tuned and various classifiers are trained;finally,a set of clinical data of bone imaging is used to evaluate the model and analyze the experimental results.The experimental results show that the positive accuracy and sensitivity reach 0.9807 and 0.9830,respectively,in the non-fused image binary classification task of the thoracic region for bone scintigraphy.(3)Image multi-classification for disease diagnosis.Considering the multi-positional features of SPECT imaging,this paper proposes a fine-grained multi-view classification method for lung cancer subtypes.First,the thoracic region is extracted from the SPECT whole-body bone scintigraphy image,and the data set is expanded;then,the multi-view classifier is customized,and the key features are extracted through the feature extraction sub-network,feature fusion sub-network and feature classification sub-network,fusion and identification;finally,model performance test and experimental result analysis are carried out.From the experimental results,the multi-view classification model constructed in this paper can not only predict the occurrence of bone metastases,but also display the subclasses of primary lung cancer,whose Acc,Pre,Roc and F-1 reach 0.7392,0.7592,0.7242 and0.7292.With the research work above,this thesis is devoted to the automatic classification technology of SPECT whole body bone imaging images,and explores the method of dichotomous and multi-segmentation of bone metastasis through data preparation and model construction.Multiple sets of experimental results show that the automatic diagnosis of bone metastases based on deep learning has certain feasibility and reliability,and can effectively identify and classify bone metastases hot spots,which is helpful for the timely diagnosis,disease staging and prognosis of patients with bone metastases.At the same time,it also shows that the deep learning method has great potential in the automatic diagnosis of various bone metastases.