Design And Implementation Of Interactive Segmentation System For Pulmonary Nodules

With the gradual use of intelligent medical imaging technology in clinical practice,intelligent assisted lung nodule screening based on interactive segmentation can relieve the work pressure of imaging doctors and change the current situation of fatigue and missed diagnosis in the repetitive and monotonous work of reading images.It can also assist doctors to provide accurate quantitative analysis,reduce the subjectivity of reading images,and improve the consistency of doctors’ diagnosis of lung nodules.It has an important reference and demonstration effect for improving the efficiency of medical services and building a precision medical service system.Under the background that the cloud PACS system will become the future development direction,after visiting two third-class hospitals in Chongqing and investigating the needs of imaging doctors on the spot,an interactive segmentation system for lung nodules based on the B/S mode was designed and implemented.Under the premise of less interaction between doctors,the experience of higher segmentation accuracy is provided,which significantly reduces the burden of segmentation of lung nodules by doctors.The main work of this thesis is as follows:(1)The overall requirement analysis,functional requirement analysis and non-functional requirement analysis of the lung nodule interactive segmentation system were carried out,and the system architecture design,system functional module and database design were completed.(2)Aiming at how to store,transmit and manage DICOM image files in the cloud,and how to parse and browse DICOM images on the browser side,by analyzing the DICOM3.0 standard and the DIOCM Web Service interface defined in the standard,using Django,Nginx,Cornerstone,and Vue.js,Element-UI and other technologies have realized the functional modules of medical image file transmission,storage,management and image reading.(3)Aiming at the problem of the imaging doctor’s heavy reading work and the time-consuming segmentation of lung nodules,the interactive segmentation algorithm of lung nodules was analyzed.In the image reading subsystem,the interactive segmentation of lung nodules based on the DEXTR model was integrated.The algorithm realizes the front-end operation interface and the back-end calling interface.(4)For the problem of online deployment of deep learning models,the online deployment methods of commonly used deep learning models are analyzed,and a back-end AI reasoning solution based on Triton Inference Server is proposed,which realizes the online deployment and management functions of AI models based on Triton..(5)Implemented the practice of Docker containerization and performance optimization for the designed and implemented interactive segmentation system of lung nodules.After a detailed system function test and performance test,the results show that the lung nodule interactive segmentation system designed and implemented in this paper can meet the needs of doctors for image reading and lung nodule segmentation,providing relevant basic practices for the cloudification of hospital information systems in the future.At the same time,it can break the data barrier and build a regional medical consortium,which provides a good solution to the plight of basic hospitals with equipment but no professional doctors.The system also provides support for labeling the attribute information of lung nodules,which is very adaptable to the rapid construction of high-quality lung nodule data sets.