Deep Learning Algorithms For Non-standard Medical Data

Intelligent medicine is highly interdisciplinarily developed based on medicine,science,en-gineering and other disciplines.It combines medical health with robotics,artificial intelligence,big data,and other research methods,and uses clinical needs as the basis for research and appli-cation.The ultimate goal is to achieve the multi-dimensional cross-integration and transforma-tion of intelligence and medicine.With the development of intelligent medicine,deep learning,as an important algorithm in the field of artificial intelligence,has made breakthroughs in many fields of medicine,making it possible for the transformation of intelligent medicine research to applications.However,medical clinical applications involve patient life and health,which are often more demanding than non-medical applications.The stability,computing performance,and accuracy of deep learning algorithms directly affect the diagnosis and treatment of patients.Therefore,deep learning algorithms for medical big data are still challenging.Different from the data collected manually for model training and verification,the data in the complex clinical environment is non-standardized,which is mainly reflected in the differences in the individual data of the patients and the diversity of the medical data of the patients at different stages of treatment.These challenges give higher requirements for the generalization ability of the learn-ing algorithm.This thesis aims to achieve clinical application and develops more robust,more stable deep learning algorithms,that can deal with the real clinical complex environment for non-standardized medical data.This thesis collects a large amount of non-standard data from West China Hospital of Sichuan University and West China Hospital of Stomatology.On this basis,it studies the prediction of postoperative complications of lung cancer for feature-missing data,the segmentation of organs at risk for label imbalanced data,and the landmark detection for the resolution-oriented unbalance data.Specifically,the main contributions of this thesis are as follows:1 The Postoperative Complication Prediction(PCP)of lung cancer was studied,and theSubtraction Recurrent Neural Networks(SRNN)based on subtraction gates and the Multi-Layer Perceptron(MLP)network based on the key feature variable extraction model,named Medi MLP,were proposed to solve the problem of missing features of electronicmedical record data.A full-cycle intelligent diagnosis and treatment platform for lungcancer was developed,which is applied to the lung cancer prognosis prediction subsys-tem.Lung cancer PCP refers to predicting the risk of postoperative complications of the patientby analyzing the observation data of the patient.Lung cancer PCP is critical to reducingthe overall mortality rate of lung cancer.Researching lung cancer PCP based on deeplearning algorithms can help patients recover fully after surgery and avoid postopera-tive complications and lung cancer recurrence.Facing the problem of missing featuresof electronic medical records,the thesis proposes SRNN and Medi MLP respectively forlung cancer PCP classification and key feature extraction tasks.The thesis collected 8459patients’ electronic medical record data from the Department of Thoracic Surgery,WestChina Hospital of Sichuan University,including 72 column characteristic variables.Theproposed models are verified on this data set.The experimental results show that SRNNhas good performance in lung cancer PCP classification experiments,and the Grad-CAMalgorithm based on the Medi MLP model has comparable performance to existing featureselection methods.Through further experimental analysis,based on the electronic medi-cal record data of West China Hospital of Sichuan University,an important conclusion isfound: the variable ”drainage tube time” is the primary key feature variable of lung cancerPCP.The thesis further develops a full-cycle intelligent diagnosis and treatment platformfor lung cancer.In the lung cancer prognosis prediction subsystem,the proposed methodcan assist clinicians in predicting the risk of complications after lung cancer surgery.Theplatform has been deployed in West China Hospital of Sichuan University,Sichuan Sec-ond Hospital of Traditional Chinese Medicine,West China Second University Hospital,Dazhou Central Hospital,and Chengdu Second People’s Hospital.2 The automatic segmentation of organs at risk was studied,and a Multi-Task Learning(MTL)model with a shared network skeleton and a Label Dependent Cross-Entroy(LDCE)loss for multi-label classification were proposed to solve the problem of pixel-level labelimbalance in CT images.The proposed method won second place in the ”Seg THOR Chal-lenge: Segmentation of THoracic Organs at Risk in CT Images” in 2019 and first placein the ”Li TS: Liver Tumor Segmentation Challenge” in March 2019 on the Coda Lab.Anintelligent segmentation system for radiotherapy target areas has been developed,whichis applied to the automatic segmentation of organs at risk.Clinical radiotherapy is the main treatment method for cancer.One of the main tasks ofradiologists is to contour organs at risk,and manual contouring is a tedious and time-consuming task.The automatic segmentation of organs at risk is an important directionfor the development of clinical radiotherapy in the future.The thesis studies the label im-balance of CT images and proposes an MTL model that shares a network skeleton.Themodel uses an encoder-decoder network to simultaneously learn two subtasks of organsegmentation and multi-label organ classification,and optimizes the False Positive Filter-ing(FPF)by using the proposed Dynamic Threshold Strategy(DTS)algorithm.The FPFalgorithm can obtain a high True Positive Rate(TPR).DTS and FPF algorithms can filterfalse positive slices in CT images with high precision.The thesis also proposes an LDCEloss function for multi-label classification,where the weight is the global conditionalprobability between two organs.When organs are highly correlated,LDCE can signifi-cantly improve the performance of multi-label organ classification tasks.The proposedmodel was verified on the real data set collected from West China Hospital of SichuanUniversity and achieved good results.It won second place in the ”Seg THOR Challenge:Segmentation of THoracic Organs at Risk in CT Images” in 2019 and first place in the”Li TS: Liver Tumor Segmentation Challenge” in March 2019 on the Coda Lab.Basedon the proposed method,an intelligent segmentation system for radiotherapy target ar-eas is developed,which is applied to the automatic segmentation of endangered organs.The system has been deployed in West China Hospital of Sichuan University,SichuanProvincial People’s Hospital,the Second People’s Hospital of Neijiang,and LiangshanPrefecture First People’s Hospital.3 The cephalometric landmark detection is studied,and a two-stage neural network archi-tecture and an end-to-end cascade learning architecture are proposed to solve the problemof resolution loss for cephalometric images.A cranio-maxillofacial intelligent cephalo-metric measurement system was developed,which realized automatic landmark detectionand gave a cephalometric measurement analysis report.Cephalometric measurement is a basic method for evaluating tooth development,treat-ment effects,and facial aesthetics.This thesis studies the resolution loss of the landmarkdetection in the cephalometric task,and proposes a two-stage neural network architec-ture.In the first stage,a Global Detection Network(GDN)is trained,which receivesthe entire image as input to generate candidate landmarks.In the second stage,the inputimage is divided into small pitches,and a Local Refine Network(LRN)is trained,whichreceives small image pitches as input to adjust the position of landmarks.LRN focuseson the detailed local features of high-resolution cephalometric images and can accuratelypredict the location of landmarks.The proposed two-stage neural network architecturetakes translation invariance as the main consideration to construct GDN and LRN,andhas achieved excellent performance on the data set collected from the West China Den-tal Hospital of Sichuan University and the public ISBI 2015 data set.In order to furtherimprove the model efficiency and resolution loss of the landmark detection,this thesisproposes an end-to-end cascaded learning architecture.By modeling the discrete approx-imation of Ordinary Differential Equation(ODE),this thesis introduces a novel cascadetraining strategy into the end-to-end Coarse-to-fine architecture,and proposes a cascade-refine model.By sharing parameters among cascaded network backbones,the model hasthe ability to continuously improve the location of landmarks,and any encoder-decodernetwork can be deployed as the network skeleton.This thesis also proposes the Landmark Label Map(LLP),which is input to the local network module as an external feature.In this way,the trained end-to-end cascaded learning architecture can generate correspond-ing landmarks for each patch.Based on the proposed method,a cranio-maxillofacial intelligent cephalometric measurement system is developed,which realizes automatic landmark detection and gives a cephalometric analysis report.The system has been de-ployed in West China Hospital of Stomatology,Dazhou Central Hospital,and the First Affiliated Hospital of Xi’an Jiaotong University.