Research And Application Of Lung Cancer Subtype Diagnosis Based On Multi-omics Data Fusion

Cancer is a serious global public health problem.Due to its rapidly increasing morbidity and mortality,lung cancer has become one of the most serious malignant tumors threatening human life and health.Lung cancer can be divided into different subtypes according to histopathological characteristics,and accurate diagnosis of lung cancer subtypes plays a crucial role in its treatment and prognosis,because different subtypes require different treatments.Although the current multi-omics data fusion method based on deep learning has significantly improved the accuracy of the diagnosis of lung cancer subtypes,due to the complex heterogeneity,high-dimensional sparsity and incompleteness of multi-omics data and other unfavorable factors,the current deep learning technology based on multi-omics data fusion still faces problems such as insufficient data fusion,over-fitting training data,and low flexibility.In addition,due to the weak interpretability of deep learning technology and the high concern from all walks of life for the protection of patient privacy,how to make the model have a good interpretability to assist the clinical diagnosis of lung cancer and ensure the privacy of patients during model training is the top priority of the current lung cancer subtype diagnosis research.In order to solve the above shortcomings,this paper models the problem of lung cancer subtype diagnosis based on the theory of multi-omics data fusion,combining Attention Mechanism,Representation Learning,Generative Adversarial Learning and Federated Learning.The main work done is as follows:1.A novel algorithm named lung cancer subtype diagnosis by fusing image-genomics data(LungDIG)is proposed to explore the impact of differences in omics feature representation and the rationality of feature fusion on subtype diagnosis.LungDIG first introduces a feature combination strategy to mitigate the impact of false positive features in image data on patient feature representation.It then uses an attention-based feature extractor to capture important mutation sites in genomic data,resulting in a richer representation of patient gene features.Finally,it utilizes an attention mechanism to more rationally fuse two omics data to achieve more accurate lung cancer subtype diagnosis.The study on the public lung cancer dataset verifies the effectiveness and advancement of the method.2.Proposed a novel method named lung cancer subtype diagnosis using weakly-paired multi-omics data(LungDWM)to achieve flexible lung cancer subtype diagnosis by balancing the shared and specific features between the weakly-paired multi-omics data.LungDWM first introduces a joint optimization strategy to balance the shared and specific information among multi-omics data,so as to make the characteristic representation of patients more abundant.It then employs generative adversarial learning to alleviate the missing omics data problem by imputing the missing omics data.Finally,it fuses real and generated data for flexible diagnosis of lung cancer subtypes.A large number of experiments have shown that LungDWM can not only diagnose lung cancer subtypes more accurately than the current state-of-the-art methods,but also have good reliability and interpretability.3.Developed a system named Federated Lung Cancer Subtype Intelligent Diagnosis System(Fed-Lung)to to alleviate problems such as poor interpretability of deep learning models and poor patient privacy protection.Based on the lightweight web development framework FLASK,the system integrates the above-mentioned LungDIG and LungDWM model architectures,and realizes the interpretable auxiliary diagnosis of lung cancer subtypes by visualizing the attention parameters in the model.Moreover,the system also introduces the federated learning framework PySyft for distributed model training to achieve a balance between patient data privacy protection and data sharing analysis.To this end,a microkernel,pluggable,high-security,and easy-to-operate intelligent diagnosis system for lung cancer subtypes was developed to accelerate the intelligent assistance process of lung cancer subtype diagnosis.