Classification Of Images For Fusion Bone Scan Using Deep Neural Architecture-based Search

"Integration of medicine and industry" is an important theme of science and technology innovation in the new era,requiring the demand side of precision medicine and the supply side of artificial intelligence to work in tandem.Nuclear medicine is a modern medical branch that integrates intelligent disease detection,diagnosis and treatment,and plays an important role in the clinical diagnosis and treatment of tumours,cardiac system and brain system diseases.However,the low spatial resolution of bone imaging in nuclear medicine severely limited the reliable analysis and accurate diagnosis of computer-aided diagnostic techniques in bone imaging computation.In order to build a reliable bone image classification model,the thesis investigated a deep neural architecture search-based fusion image classification method,which mainly includes disease diagnosis-oriented bone image two-class classification and disease typing-oriented fusioned bone image fusion multi-class classification.In summary,the following research work is conducted in this thesis:(1)Disease diagnosis-oriented bone image two-class classification.Targeting at automated detection of skeletal metastasis caused by lung cancer,this thesis proposes a two-class classification model of bone imaging images for automatic diagnosis of metastases from tumours.In particular,it is split into two parts: first,building a fivestage search space based on the idea of Res Net model structure,in which it contains chained and Cell structures.Second,the anterior-posterior feature information was fused using the feature fusion method.In this thesis,evaluation results show that the accuracy of the proposed two-class classification model for bone images for disease diagnosis reaches 0.7705 and 0.9205 in two sets of clinical data,which exceeds that of current artificially designed neural networks.(2)Disease typing-oriented fusioned bone image fusion multi-class classification.Targeting the problem of similar characteristics of tumour subclasses of bone metastases and the difficulty of classification,we propose a multi-class classification model of fusion images of bone imaging for tumour subclass recognition.In particular,it is split into two parts: first,basing on the anterior-posterior view of the scintigraphic images,the method of pixel-level fusion of anterior and posterior views is proposed.Firstly,the bone imaging image is extracted from the human contour information.Then the four correction points of the image are identified based on the contour information.And pixel-level summation fusion of anterior and posterior bit views was performed according to the correction points.Second,building a multi-class classification model for fusion images of bone images for tumor subclass recognition.Two models were constructed,including a deep neural architecture search model based on early stopping and a traditional deep neural network.The experimental results showed that the accuracy and F-1 of the triple classification reached 0.7852 and 0.7823 respectively,verifying the effectiveness of the multi-class classification model of fusion images of bone imaging for tumour subclass recognition.Through the above research work,we have focused on the automatic classification technology of nuclear medicine bone imaging images,and explored the methodologies of dichotomous and multi-classification of bone metastases through neural architecture search.The results of several experiments show that the automatic diagnosis of fused bone images based on deep neural architecture search is feasible and reliable,which can help clinical assistants in the timely diagnosis,disease staging and prognosis of bone metastases and fill the gap in the field,and demonstrate the great potential and value of deep learning technology in the field of computational nuclear medicine.