Research On Growth Prediction Of Early Lung Cancer Via Follow-up Imaging

Lung cancer is the cancer with the highest incidence and mortality in the world,and seriously threatens human health and safety.The five-year survival rate for lung cancer is only about 15%,but for early stage lung cancer after standardized treatment,the five-year expected survival rate is close to 90%.The screening of imaging technology can effectively detect early lung cancer and is one of the important means to reducethe mortality rate of lung cancer patients.Low-dose spiral CT is an internationally recognized imaging method for improving the detection rate of early lung cancer.The use of CT images to compare lung cancer across periods can effectively detect early lung cancer and reveal the growth pattern of tumors,which contribute to the early classification and staging of pulmonary nodules,and is of great significance for the formulation of corresponding treatment plans.However,the current computer-assisted intelligent diagnosis are mostly aimed at image slices in independent periods,ignored the research on the characteristics and development trends of lesions that gradually evolve over time.This dissertation starts with medical image processing methods,mainly for intelligent screening and auxiliary diagnosis of lung cancer.After penetratingly exchanges with partner hospitals,this dissertation focuses on lung cancer CT images,combined with multi-modal data analysis of PET images and machine learning related technologies to explore tumor segmentation in CT image sequences,and construct tumor malignancy classification and staging classification model,further study the longitudinal prediction to achieve early warning of lung cancer.In particular,the main research work and achievements include the following points.(1)Aiming at the problem that the existing pulmonary nodule segmentation methods are designed to segment a single type of pulmonary nodule and lack the universality of nodule segmentation,this thesis proposes a general nodule segmentation method based on the Dual Attention Segmentation Adversarial Network model.The method mainly includes two parts:segmentation network module and discriminative network module.Among them,the segmentation network module will first segment the pulmonary nodules in the CT image,and the discriminative network module will determine the difference between the segmented pulmonary nodules and the gold truth according to the distance between the two as the performance of the segmentation network.The index is fed back to the segmentation network,and then the parameters of the segmentation network are adjusted.The two are iteratively trained against each other,and finally the pulmonary nodules are segmented.After obtaining the results of segmentation of nodules,this article uses visual attention and spatial attention to improve the segmentation results.Among them,visual attention mainly extracts the semantic features of pulmonary nodules,while spatial attention is to extract spatial characteristics of pulmonary nodules in CT sequences.Finally,the accuracy of segmentation pixels reached90.14%.The experimental results show that the method in this thesis can effectively segment pulmonary nodules from CT images.The segmentation adversarial network after adding dual attention can improve the segmentation effect of pulmonary nodules.(2)Aiming at the problems of existing benign and malignant pulmonary nodule recognition algorithms,such as high accuracy and complexity,and strong inexplicability,this thesis designs a pulmonary nodule malignancy classification algorithm based on weighted ensemble classification.This method is different from the diagnosis methods of pulmonary nodules which directly classifying benign and malignant lung nodules,it classifies the malignant degree of pulmonary nodules,using Denoising Auto-encoders,Residual networks,and gray-level co-occurrence matrices and geometric parameters to characterize the characteristics of pulmonary nodules.Different classifiers are used to classify the extracted features,and the weight of the classifier in the integrated model is dynamically adjusted according to the error rate of the classifier,and finally ensemble into an automatic diagnosis model of lung nodules.In order to verify the effectiveness of the method,this thesis verifies the images of pulmonary nodules extracted from the collected data sets.Through comparative experiments,ablation experiments within the method and the selection of different models,the accuracy reached 93.10%,the accuracy reached 83.85%,and the sensitivity reached 81.75%,which comprehensively verify that the integration method proposed in this article is effective(3)Aiming at the different expressions of tumor characteristics in PET and CT singlemodality images,as well as the insufficient diagnosis of lung cancer staging,this thesis designs a PET-CT image fusion method based on Siamese Auto-encoder,and to classify lung cancer staging.This method first decomposes the original PET image and CT image into two different sub-bands,low-frequency and high-frequency.The high-frequency sub-band mainly reflects the salient characteristics of the image and the characteristics of the brightness mutation,which is mainly manifested in the details of the edge of the object in the image,the boundary of the region and other details.This article uses Siamese Auto-encoder for fusion;the low-frequency sub-band concentrates the main energy of the image,which mainly reflects the background and average characteristics of the image.This thesis uses principal component analysis to fuse the low-frequency sub-band.In this thesis,reference and non-reference objective evaluation indicators are selected to evaluate the fusion results.The quantitative indicators after fusion and the qualitative comparison with other methods show that the method proposed in this thesis can effectively fuse medical images of the two modalities,and has better performance in the presentation of details.(4)Aiming at the fact that single-point time images cannot fully characterize the development process of lung cancer,this thesis proposes a tumor longitudinal prediction method based on the image sequence encoding and decoding model.By extracting the benign and malignant features of early lung tumors and the tumor staging features,the encodingdecoding structure maps the high-dimensional and complex tumor sequence CT images into the shallow feature space,uses the superficially representable tumor features to represent the original tumor for prediction,and effectively reconstruct the predicted tumor back to the original image.Through the conditional recurrent unit and the flow-based model,the prospective prediction and retrospective prediction of tumor characteristics are realized respectively,so as to restore the tumors at various stages,complete the modeling of the tumor growth and evolution law across periods,and explore the evolution of tumor serial images in the time dimension.Experiments with follow-up data collected in cooperative hospitals show that the proposed method can effectively realize the longitudinal prediction of tumors.Finally,this dissertation summarizes the work done,looks forward to the direction of future work,and discusses further research plans.