Research On Intelligent Lung Tumor Analysis Technology Based On Multimodal Data Fusion

Cancer threatens human health worldwide.Among the world's cancer patients,lung cancer incidence rate is second(males only after prostate cancer,females only second to breast cancer).The highest mortality rate is cancer,which is of great significance for early diagnosis.In clinical,imaging diagnosis information has high clinical value for early diagnosis and treatment of lung cancer.At present,most of the computer-aided research of lung cancer is carried out with the aid of imaging information.With the rapid development of medical technology,the types and quantity of medical data are constantly enriched and increased.Because of the complexity of lung tumor diagnosis,it is very important to help doctors to grasp the information describing different signs and examinations of patients with computer-aided technology.The method of using multiple different types of data(also called multimodal data)has become an important driving force of intelligent medical system.Its application research scope has been covered in all aspects of screening discovery,disease diagnosis and treatment prognosis evaluation.In this treatise,the following research work is done by using artificial intelligence technology and multimodal medical data to make lung tumor diagnosis as the breakthrough point.(1)Research on distance metric learning based on Gaussian mixture in data understanding and model generalization.Clinical medical data is different from other general data.In contrast,most of them are scattered in different hospitals and institutions in the form of small samples.When using artificial intelligence technology based on big data to study clinical problems,due to the limitation of the sample size,the training model with small samples is unstable and prone to bias,and the generalization performance is questionable.From the perspective of data understanding,this article proposes a method of the distance metric learning based on Gaussian mixture(DML-GMM),which can help us to effectively understand the distribution of data and grasp the essential characteristics of data.Under the condition of understanding the characteristics of the data,the process of artificial intelligence model training is optimized to reduce the bias and variance of the model.At the same time,it can help us find singular samples in practice,help us to update the model iteratively,and understand the scope of application and generalization performance of the model.The data understanding method proposed in this work is an important foundation and guarantee for the follow-up work under the condition of small clinical samples.(2)Research on the reliability of deep learning models: confidence score based on distanceDeep learning technology is like a "black box".Although it is effective in some applications,it lacks a certain degree of interpretability.However,in the medical field applications,it usually has certain requirements for its reliability and interpretability.In this context,in response to this demand,we have studied the confidence of the model,and proposed a confidence score based on distance to measure the confidence of the classification model.This confidence score is used to explain the prediction results.Reliability.The confidence score can be used to quantify and evaluate the influence of different modal information on the confidence of the model when analyzing the multimodal fusion classification model,and explain the role and influence of different types of data on the model from the perspective of confidence.(3)Research on identification technology of pathological subtypes of lung adenocarcinoma based on multimodal dataThe pathological non-invasive differentiation of lung cancer is a very challenging problem in clinical diagnosis.This study collected data from 1946 patients with clinically confirmed lung adenocarcinoma,and used deep learning technology to combine various examination information commonly used in the diagnosis process:patient general clinical information,serum tumor markers and CT imaging examinations,and proposed a The multi-modal fusion lung adenocarcinoma pathological classification model based on the attention mechanism is used to distinguish the invasive and non-invasive lung adenocarcinoma.The method achieved88.4% accuracy and 0.958 AUC on the test data,which is better than the current one.There are related studies.(4)Prediction of the efficacy of radiofrequency ablation for lung tumors based on the fusion of deep learning and radiomics.In this applied study of clinical postoperative evaluation(prediction of curative effect after radiofrequency ablation of lung tumor),a prediction model of curative effect after radiofrequency ablation of lung tumor based on deep learning and radiomics fusion is proposed to evaluate the prognosis and curative effect.The model integrates the multimodal image information of lung tumor before and after radiofrequency ablation.In the case of small sample,the better prediction performance is obtained.