Artificial Intelligence in Medicine : Technical Basis and Clinical Applications

Artificial Intelligence Medicine: Technical Basis and Clinical Applications presents a comprehensive overview of the field, ranging from its history and technical foundations, to specific clinical applications and finally to prospects. Artificial Intelligence (AI) is expanding across all domains at a breakneck speed. Medicine, with the availability of large multidimensional datasets, lends itself to strong potential advancement with the appropriate harnessing of AI. The integration of AI can occur throughout the continuum of medicine: from basic laboratory discovery to clinical application and healthcare delivery. Integrating AI within medicine has been met with both excitement and scepticism. By understanding how AI works, and developing an appreciation for both limitations and strengths, clinicians can harness its computational power to streamline workflow and improve patient care. It also provides the opportunity to improve upon research methodologies beyond what is currently available using traditional statistical approaches. On the other hand, computers scientists and data analysts can provide solutions, but often lack easy access to clinical insight that may help focus their efforts. This book provides vital background knowledge to help bring these two groups together, and to engage in more streamlined dialogue to yield productive collaborative solutions in the field of medicine. Provides history and overview of artificial intelligence, as narrated by pioneers in the field Discusses broad and deep background and updates on recent advances in both medicine and artificial intelligence that enabled the application of artificial intelligence Addresses the ever-expanding application of this novel technology and discusses some of the unique challenges associated with such an approach 人工智能医学:技术基础和临床应用全面概述了该领域,从其历史和技术基础,到具体的临床应用,最后到前景。人工智能 (AI) 正在以惊人的速度扩展到所有领域。医学凭借大型多维数据集的可用性,可以通过适当利用人工智能实现强大的潜在进步。人工智能的整合可以发生在整个医学的连续过程中:从基本的实验室发现到临床应用和医疗保健服务。将人工智能整合到医学中既令人兴奋又令人怀疑。通过了解 Artificial Intelligence in Medicine Copyright Dedication Contents List of contributors Foreword References Preface Acknowledgments 1 Artificial intelligence in medicine: past, present, and future 1.1 Introduction 1.2 A brief history of artificial intelligence and its applications in medicine 1.3 How intelligent is artificial intelligence? 1.4 Artificial intelligence, machine learning, and precision medicine 1.5 Algorithms and models 1.6 Health data sources and types 1.7 The promise 1.8 The challenges 1.8.1 Quality and completeness of training data 1.8.2 Trust and performance: the case for model interpretability 1.8.3 Beyond performance and interpretability: causality 1.8.4 Defining the question, measuring real-world impact 1.8.5 Maximizing information gain across modalities, tasks, populations, and time 1.8.6 Quality assessment and expert supervision 1.9 Making it a reality: integrating artificial intelligence into the human workforce of a learning health system References 2 Artificial intelligence in medicine: Technical basis and clinical applications 2.1 Introduction 2.2 Technology used in clinical artificial intelligence tools 2.2.1 Elements of artificial intelligence algorithms 2.2.1.1 Activation functions 2.2.1.2 Fully connected layer 2.2.1.3 Dropout 2.2.1.4 Residual blocks 2.2.1.5 Initialization 2.2.1.6 Convolution and transposed convolution 2.2.1.7 Inception layers 2.2.2 Popular artificial intelligence software architectures 2.2.2.1 Neural networks and fully connected networks 2.2.2.2 Convolutional neural networks 2.2.2.3 U-Nets and V-Nets 2.2.2.4 DenseNets 2.2.2.5 Generative adversarial networks 2.2.2.6 Hybrid generative adversarial network designs 2.3 Clinical applications 2.3.1 Applications of regression 2.3.1.1 Bone age 2.3.1.2 Brain age 2.3.2 Applications of segmentation 2.3.3 Applications of classification 2.3.3.1 Detection of disease 2.3.3.2 Diagnosis of disease class 2.3.3.3 Prediction of molecular markers 2.3.3.4 Prediction of outcome and survival 2.3.4 Deep learning for improved image reconstruction 2.4 Future directions 2.4.1 Understanding what artificial intelligence “sees” 2.4.2 Workflow 2.5 Conclusion References 3 Deep learning for biomedical videos: perspective and recommendations 3.1 Introduction 3.2 Video datasets 3.3 Semantic segmentation 3.4 Object detection and tracking 3.5 Motion classification 3.6 Future directions and conclusion References 4 Biomedical imaging and analysis through deep learning 4.1 Introduction 4.2 Tomographic image reconstruction 4.2.1 Foundation 4.2.2 Computed tomography 4.2.3 Magnetic resonance imaging 4.2.4 Other imaging modalities 4.3 Image segmentation 4.3.1 Introduction 4.3.2 Localization versus segmentation 4.3.3 Fully convolutional networks 4.3.4 Regions with convolutional neural network features 4.3.5 A priori information 4.3.6 Manual labeling 4.3.7 Semisupervised and unsupervised approaches 4.4 Image registration 4.4.1 Single-modality image registration 4.4.2 Multimodality image registration 4.5 Deep-learning-based radiomics 4.5.1 Detection 4.5.2 Characterization and diagnosis 4.5.3 Prognosis 4.5.4 Assessment and prediction of response to treatment 4.5.5 Assessment of risk of future cancer 4.6 Summary and outlook References 5 Expert systems in medicine 5.1 Introduction 5.2 A brief history 5.3 Methods 5.3.1 Expert system architecture 5.3.2 Knowledge representation and management 5.3.3 Uncertainty, probabilistic reasoning, fuzzy logic 5.3.3.1 Uncertainty 5.3.3.2 Probabilistic reasoning 5.3.3.3 Fuzzy logic 5.4 Applications 5.4.1 Computer-assisted diagnosis 5.4.2 Computer-assisted therapy 5.4.3 Medication alert systems 5.4.4 Reminder systems 5.5 Challenges 5.5.1 Workflow integration 5.5.2 Clinician acceptance and alert fatigue 5.5.3 Knowledge maintenance 5.5.4 Standard, transferability, and interoperability 5.6 Future directions References 6 Privacy-preserving collaborative deep learning methods for multiinstitutional training without sharing patient data 6.1 Introduction 6.2 Variants of distributed learning 6.2.1 Model ensembling 6.2.2 Cyclical weight transfer 6.2.3 Federated learning 6.2.4 Split learning 6.3 Handling data heterogeneity 6.4 Protecting patient privacy 6.5 Publicly available software 6.6 Conclusion References 7 Analytics methods and tools for integration of biomedical data in medicine 7.1 The rise of multimodal data in biology and medicine 7.1.1 The emergence of various sequencing techniques 7.1.1.1 Bulk sequencing 7.1.1.2 Single-cell sequencing 7.1.2 The increasing need for combining images and omics in clinical applications 7.1.2.1 Various modalities of images in clinics 7.1.2.2 The rise of radiomics: combine medical images with omics 7.1.3 The availability of large-scale public health data 7.2 The challenges in multimodal data—problems with learning from multiple sources of data 7.2.1 The imperfect generation of single-cell data 7.2.1.1 The complementariness of various sources of data 7.2.2 The issues of generalizability of machine learning 7.3 Machine learning algorithms in integrating medical and biological data 7.3.1 Genome-wide data integration with machine learning 7.3.1.1 How to integrate various omics for cancer subtyping 7.3.1.2 How to integrate single-cell multiomics for precision medicine 7.3.2 Data integration beyond omics—an example with cardiovascular diseases 7.3.2.1 How to integrate various image modalities such as magnetic resonance imaging computed tomography scans 7.3.2.2 How to better the diagnosis by linking images with electrocardiograms 7.3.3 Multimodal decision-making in clinical settings 7.4 Future directions References 8 Electronic health record data mining for artificial intelligence healthcare 8.1 Introduction 8.2 Overview of the electronic health record 8.2.1 History of the electronic health record 8.2.2 Core functions of an electronic health record 8.2.3 Electronic health record ontologies and data standards 8.3 Clinical decision support 8.3.1 Healthcare primed for clinical decision support 8.4 Areas of artificial intelligence augmentation for electronic health records 8.4.1 Artificial intelligence to improve data entry and extraction 8.4.2 Optimizing care 8.4.3 Predictions 8.4.4 Hospital outcomes 8.4.5 Sepsis and infections 8.4.6 Oncology 8.5 Limitations of artificial intelligence and next steps References 9 Roles of artificial intelligence in wellness, healthy living, and healthy status sensing 9.1 Introduction 9.2 Diet 9.3 Fitness and physical activity 9.4 Sleep 9.5 Sexual and reproductive health 9.6 Mental health 9.7 Behavioral factors 9.8 Environmental and social determinants of health 9.9 Remote screening tools 9.10 Conclusion References 10 The growing significance of smartphone apps in data-driven clinical decision-making: Challenges and pitfalls 10.1 Introduction 10.2 Distribution of apps in the field of medicine 10.3 Distribution of apps over different locations 10.4 Reporting applications development approaches 10.5 Decision-support modalities 10.6 Camera-based apps 10.7 Guideline/algorithm applications 10.8 Predictive modeling applications 10.9 Sensor-linked apps 10.10 Discussion 10.11 Summary References 11 Artificial intelligence for pathology 11.1 Introduction 11.2 Deep neural networks 11.2.1 Convolutional neural networks 11.2.2 Fully convolutional networks 11.2.3 Generative adversarial networks 11.2.4 Stacked autoencoders 11.2.5 Recurrent neural networks 11.3 Deep learning in pathological image analysis 11.3.1 Image classification 11.3.1.1 Image-level classification 11.3.1.2 Object-level classification 11.3.2 Object detection 11.3.2.1 Detection of particular types of objects 11.3.2.2 Detection of objects without category labeling 11.3.2.3 Detection of objects with category labeling 11.3.3 Image segmentation 11.3.3.1 Nucleus/cell segmentation 11.3.3.2 Gland segmentation 11.3.3.3 Segmentation of other biological structures or tissues 11.3.4 Stain normalization 11.3.5 Image superresolution 11.3.6 Computer-aided diagnosis 11.3.7 Others 11.4 Summary 11.4.1 Open challenges and future directions of deep learning in pathology image analysis 11.4.1.1 Quality control 11.4.1.2 High image dimension 11.4.1.3 Object crowding 11.4.1.4 Data annotation issues 11.4.1.5 Integration of different types of input data 11.4.2 Outlook of clinical adoption of artificial intelligence 11.4.2.1 Potential applications 11.4.2.2 Barriers to clinical adoption 11.4.2.2.1 Lagging adoption of digital pathology 11.4.2.2.2 Lack of standards for interfacing AI to clinical systems 11.4.2.2.3 Regulatory concerns 11.4.2.2.4 Computational requirements 11.4.2.2.5 Algorithm explainability 11.4.2.2.6 Pathologists’ skepticism References 12 The potential of deep learning for gastrointestinal endoscopy—a disruptive new technology 12.1 Introduction 12.2 Applications of artificial intelligence in video capsule endoscopy 12.2.1 Introduction 12.2.2 Decreasing read time 12.2.3 Anatomical landmark identification 12.2.4 Improving sensitivity 12.2.5 Recent developments 12.3 Applications of artificial intelligence in upper endoscopy 12.3.1 Introduction 12.3.2 Esophageal cancer 12.3.3 Gastric cancer 12.3.4 Upper endoscopy quality 12.3.5 Future directions 12.4 Applications of artificial intelligence in colonoscopy 12.4.1 Introduction 12.4.2 Cecal intubation rate and cecal intubation time 12.4.3 Withdrawal time 12.4.4 Boston Bowel Prep Scoring 12.4.5 Polyp detection 12.4.6 Polyp size 12.4.7 Polyp morphology 12.4.8 Polyp pathology 12.4.9 Tools 12.4.10 Mayo endoscopic subscore 12.5 Conclusion 12.6 Future directions References 13 Lessons learnt from harnessing deep learning for real-world clinical applications in ophthalmology: detecting diabetic r... 13.1 Introduction 13.2 Historical artificial intelligence for diabetic retinopathy 13.3 Deep learning era 13.4 Lessons from interpreting and evaluating studies 13.5 Important factors for real-world usage 13.6 Regulatory approvals and further validation 13.7 Toward patient impact and beyond 13.8 Summary Conflict of interest References 14 Artificial intelligence in radiology 14.1 Introduction 14.2 Thoracic applications 14.2.1 Pulmonary analysis in chest X-ray 14.2.2 Pulmonary analysis in computerized tomography 14.2.2.1 Lung, lobe, and airway segmentation 14.2.2.2 Interstitial lung disease pattern recognition 14.3 Abdominal applications 14.3.1 Pancreatic cancer analysis in computerized tomography and magnetic resonance imaging 14.3.1.1 Pancreas segmentation in computerized tomography and magnetic resonance imaging 14.3.1.2 Pancreatic tumor segmentation and detection in computerized tomography and magnetic resonance imaging 14.3.1.3 Prediction and prognosis with pancreatic cancer imaging 14.3.2 AI in other abdominal imaging 14.4 Pelvic applications 14.5 Universal lesion analysis 14.5.1 DeepLesion dataset 14.5.2 Lesion detection and classification 14.5.3 Lesion segmentation and quantification 14.5.4 Lesion retrieval and mining 14.6 Conclusion References 15 Artificial intelligence and interpretations in breast cancer imaging 15.1 Introduction 15.2 Artificial intelligence in decision support 15.3 Artificial intelligence in breast cancer screening 15.4 Artificial intelligence in breast cancer risk assessment: density and parenchymal pattern 15.5 Artificial intelligence in breast cancer diagnosis and prognosis 15.6 Artificial intelligence for treatment response, risk of recurrence, and cancer discovery 15.7 Conclusion and discussion References 16 Prospect and adversity of artificial intelligence in urology 16.1 Introduction 16.2 Basic examinations in urology 16.2.1 Urinalysis and urine cytology 16.2.2 Ultrasound examination 16.3 Urological endoscopy 16.3.1 Cystoscopy and transurethral resection of the bladder 16.3.2 Ureterorenoscopy 16.4 Andrology 16.5 Diagnostic imaging 16.5.1 Prostate 16.5.2 Kidney 16.5.3 Ureter and bladder 16.6 Robotic surgery 16.6.1 Preoperative preparation 16.6.2 Navigation 16.6.3 Automated maneuver 16.7 Risk prediction 16.8 Future direction References 17 Meaningful incorporation of artificial intelligence for personalized patient management during cancer: Quantitative imag... 17.1 Introduction 17.1.1 Workflow 17.1.1.1 Data acquisition 17.1.1.2 Preprocessing 17.1.1.3 Model building and evaluation 17.1.1.4 Inference 17.1.2 Meaningful incorporation of machine learning 17.2 Quantitative imaging 17.2.1 Brief overview of the physics of imaging modalities 17.2.2 Use of artificial intelligence in different stages of a quantitative imaging workflow 17.3 Risk assessment in cancer 17.4 Therapeutic outcome prediction 17.4.1 Chemotherapy 17.4.2 Radiation therapy 17.5 Using artificial intelligence meaningfully 17.6 Summary References 18 Artificial intelligence in oncology Abbreviations 18.1 Introduction 18.2 Electronic health records and clinical data warehouse 18.2.1 Data reuse for research purposes 18.2.2 Data reuse and artificial intelligence 18.2.3 Data reuse for patient care 18.3 Artificial intelligence applications for imaging in oncology 18.3.1 Applications in oncology for diagnosis and prediction 18.3.1.1 Computer vision and image analysis 18.3.1.2 Radiomics: data-driven biomarker discovery 18.3.1.3 Artificial intelligence–assisted diagnosis and monitoring in oncology 18.3.1.4 Treatment outcome assessment and prediction 18.3.2 Applications in oncology to improve exam quality and workflow 18.3.2.1 Improvement of image acquisition 18.3.2.2 Image segmentation 18.3.2.3 Improved workflow 18.3.2.4 Interventional radiology 18.4 Artificial intelligence applications for radiation oncology 18.4.1 Treatment planning 18.4.1.1 Segmentation 18.4.1.1.1 Brain 18.4.1.1.2 Head and neck 18.4.1.1.3 Lung 18.4.1.1.4 Abdomen 18.4.1.1.5 Pelvis 18.4.1.2 Dosimetry 18.4.2 Outcome prediction 18.4.2.1 Treatment response 18.4.2.1.1 Brain 18.4.2.1.2 Head and neck 18.4.2.1.3 Lung 18.4.2.1.4 Esophagus 18.4.2.1.5 Rectum 18.4.2.2 Toxicity 18.5 Future directions References 19 Artificial intelligence in cardiovascular imaging 19.1 Introduction 19.2 Types of machine learning 19.3 Deep learning 19.4 Role of artificial intelligence in echocardiography 19.5 Role of artificial intelligence computed tomography 19.6 Role of artificial intelligence in nuclear cardiology 19.7 Role of artificial intelligence in cardiac magnetic resonance imaging 19.8 Role of artificial intelligence in electrocardiogram 19.9 The role of artificial intelligence in large databases 19.10 Our views on machine learning 19.11 Conclusion References 20 Artificial intelligence as applied to clinical neurological conditions 20.1 Introduction to artificial intelligence in neurology 20.2 Integration with clinical workflow 20.2.1 Diagnosis 20.2.2 Risk prognostication 20.2.3 Surgical planning 20.2.4 Intraoperative guidance and enhancement 20.2.5 Neurophysiological monitoring 20.2.6 Clinical decision support 20.2.7 Theoretical neurological artificial intelligence research 20.3 Currently adopted methods in clinical use 20.4 Challenges 20.4.1 Data volume 20.4.2 Data quality 20.4.3 Generalizability 20.4.4 Interpretability 20.4.5 Legal 20.4.6 Ethical 20.5 Conclusion References 21 Harnessing the potential of artificial neural networks for pediatric patient management 21.1 Introduction 21.2 Applications of artificial intelligence in diagnosis and prognosis 21.2.1 Prematurity 21.2.2 Childhood brain tumors 21.2.3 Epilepsy and seizure disorders 21.2.4 Autism spectrum disorder 21.2.5 Mood disorders and psychoses 21.2.6 Hydrocephalus 21.2.7 Traumatic brain injury 21.2.8 Molecular mechanisms of disease 21.2.9 Other disease entities 21.3 Transition to treatment decision-making using artificial intelligence 21.4 Future directions References 22 Artificial intelligence–enabled public health surveillance—from local detection to global epidemic monitoring and control 22.1 Introduction 22.2 Artificial intelligence–enhanced data analysis for outbreak detection and early warning 22.2.1 Analyzing data collected from the physical world 22.2.2 Analyzing data from the cyberspace 22.2.3 From syndromic to pre-syndromic disease surveillance: A safety net for public health 22.3 Artificial intelligence–enhanced prediction in support of public health surveillance 22.3.1 Time series prediction based on dependent variables 22.3.2 Time series prediction based on dependent and independent variables 22.4 Artificial intelligence–based infectious disease transmission modeling and response assessment 22.4.1 Modeling disease transmission dynamics based on machine learning and complex networks 22.4.2 Modeling disease transmission dynamics based on multiagent modeling 22.5 Internet-based surveillance systems for global epidemic monitoring 22.6 Conclusion References 23 Regulatory, social, ethical, and legal issues of artificial intelligence in medicine 23.1 Introduction 23.2 Ethical issues in data acquisition 23.2.1 Ethical issues arising from each type of data source 23.2.1.1 Ethical issues common to all data sources: Privacy and confidentiality 23.2.1.2 Ethical issues unique to each data source: Issues of consent 23.2.1.2.1 Issues of consent with data from research repositories 23.2.1.2.2 Return of results from research repositories 23.2.1.2.3 Issues of consent with clinical or public health data 23.2.1.2.4 Incidental or secondary findings in clinical or public health data 23.2.1.2.5 Issues of consent with nonclinically collected data 23.2.2 Future directions: Toward a new model of data stewardship 23.3 Application problems: Problems with learning from the data 23.3.1 Values embedded in algorithm design 23.3.2 Biases in the data themselves 23.3.3 Biases in the society in which the data occurs 23.3.4 Issues of implementation 23.3.5 Summary 23.4 Issues in regulation 23.4.1 Challenges to existing regulatory frameworks 23.4.2 Challenges in oversight and regulation of artificial intelligence used in healthcare 23.4.3 Regulation of safety and efficacy 23.4.4 Privacy and data protection 23.4.5 Transparency, liability, responsibility, and trust 23.5 Implications for the ethos of medicine 23.6 Future directions References 24 Industry perspectives and commercial opportunities of artificial intelligence in medicine 24.1 Introduction 24.2 Exciting growth of artificial intelligence in medicine 24.3 A framework on development of artificial intelligence in medicine 24.3.1 The power of public attention and funding 24.3.2 Technology relies on continuous innovation 24.3.3 Practical applications bring the innovation to the real world 24.3.4 Market adoption defines the success 24.3.5 Apply the framework to the current and future market 24.3.6 Patient privacy 24.3.7 Approving a moving target 24.3.8 Accountability and transparency 24.4 Business opportunity of artificial intelligence in medicine References 25 Outlook of the future landscape of artificial intelligence in medicine and new challenges 25.1 Overview of artificial intelligence in health care 25.1.1 Models dealing with input and output data from the same domain 25.1.2 Deep learning as applied to problems with input and output related by physical/mathematical law 25.1.3 Models with input and output data domains related by empirical evidence or measurements 25.1.4 Applications beyond traditional indications 25.2 Challenges ahead and issues relevant to the practical implementation of artificial intelligence in medicine 25.2.1 Technical challenges 25.2.2 Data, data curation, and sharing 25.2.3 Data and potential bias in artificial intelligence 25.2.4 Workflow and practical implementation 25.2.5 Clinical tests 25.2.6 Economical, political, social, ethical, and legal aspects 25.2.7 Education and training 25.3 Future directions and opportunities 25.4 Summary and outlook References Index