데이터 유형에 따른 AI 기반 예측 및 분석 자동화 연구 = Data Type Oriented AI Methods for Automated Prediction and Analysis

다국어 초록 (Multilingual Abstract)

Despite the rapid growth of experimental and imaging data, conventional empirical and trial-and-error–based research approaches continue to face fundamental limitations in terms of efficiency and reproducibility. These limitations arise primarily from constraints in data scale, complex nonlinear interactions among variables, and the inherent subjectivity of human interpretation. In particular, research problems involving limited experimental datasets combined with high-dimensional and unstructured data remain challenging, as traditional analytical techniques often fail to provide reliable predictions or robust quantitative interpretations.
To address these challenges, this dissertation proposes an integrated artificial intelligence (AI)–based analytical framework that systematically incorporates both structured and unstructured data. The proposed framework is applied to a diverse set of experimental- and image-based research scenarios, and its effectiveness is rigorously validated through quantitative evaluation and experimental verification.
In the first part of this study, machine learning models based on multi-output regression are developed to predict key performance indicators using limited-scale structured experimental data. Ensemble-based learning algorithms, coupled with automated hyperparameter optimization, are employed to effectively capture nonlinear interactions among input variables. The predicted optimal conditions exhibit strong agreement with experimental results, demonstrating the reliability of the proposed modeling approach. Furthermore, by leveraging open datasets, this study conducts a comparative analysis between conventional manually coded machine learning workflows and no-code generative AI–based modeling approaches. The results reveal that generative AI significantly improves accessibility and efficiency in model construction; however, expert-driven validation remains essential to ensure scientific interpretability and reliability.
In the second part, deep learning–based automated analysis techniques are applied to unstructured image and video data. In the context of agricultural pest and disease diagnosis, model performance is substantially improved through large-scale dataset restructuring, mitigation of data imbalance, and systematic data augmentation. These results highlight the critical role of a data-centric approach in enhancing model accuracy. In addition, a deep learning–based Computer-Assisted Sperm Analysis (CASA) system is developed for analyzing sperm motility under high-density microscopy conditions. By integrating object detection and multi-object tracking models, the proposed system enables real-time, stable tracking of individual sperm cells and automatically computes motility parameters in accordance with World Health Organization (WHO) guidelines. Compared with conventional manual analysis, the system achieves markedly improved accuracy and processing speed.
Overall, this dissertation demonstrates that artificial intelligence can serve not merely as an auxiliary analytical tool, but as a core research methodology encompassing experimental design, data interpretation, and quantitative evaluation. By providing a unified framework for the integrated analysis of structured and unstructured data, this work presents a new research paradigm that overcomes limitations related to data scale and modality. The proposed approach is expected to be broadly applicable across scientific, engineering, and biomedical domains, offering a versatile and extensible AI-driven framework for future data-intensive research.

목차 (Table of Contents)

• - 목 차 - i
• LIST OF FIGURES viii
• LIST OF TABLES ix
• ABSTRACT x
• 1. Introduction 1
• 1.1 Data-driven research as a paradigm shift in science and engineering 1
• 1.2 Challenges of variability and reproducibility in experimental data 3
• 1.3 Machine learning for structured experimental data 4
• 1.4 Emergence of generative AI and no-code ML workflows 5
• 1.5 Deep learning for unstructured image and video data 6
• 1.6 Scope, novelty, and organization of this dissertation 7
• 1.7 Background 8
• 1.7.1 Data-driven science and the evolution of research methodologies 8
• 1.7.2 Characteristics of structured and unstructured scientific data 9
• 1.7.3 Machine learning for structured experimental datasets 10
• 1.7.4 Generative AI and no-code machine learning paradigms 11
• 1.7.5 Deep learning for image and video-based scientific analysis 12
• 1.7.6 Need for an integrated AI framework across data modalities 13
• 2. Methods 14
• 2. 1. Experimental-data-driven prediction of PSC efficiency 14
• 2.1.1 Formulation of the Regression Problem 14
• 2.1.2 Dataset Preparation and Preprocessing 14
• 2.1.3 Machine Learning Model Selection 15
• 2.1.4 Hyperparameter Optimization 15
• 2.1.5 Model Implementation and Reproducibility 16
• 2. 2. Open-data-driven prediction of PSC efficiency 17
• 2.2.1 Data preparation and dataset construction 17
• 2.2.2 Machine learning model evaluation using PyCaret 18
• 2.2.3 ML model selection and training strategy 19
• 2.2.4 One-hot encoding of categorical device structures 19
• 2.2.5 Model interpretability and validation using SHAP analysis 20
• 2. 3. Predicting PSC Efficiency Using Generative AI 21
• 2.3.1 Human-coded machine learning workflow 21
• 2.3.2 Generative AIbased no-code machine learning workflow 21
• 2.3.3 Model interpretability and validation 22
• 2.3.4 Summary of methodological contribution 23
• 2. 4. Image datadriven pest and disease diagnosis system 24
• 2.4.1 Construction of the initial image dataset (V1) and baseline model evaluation 24
• 2.4.2 Quantitative analysis of data imbalance and identification of additional data requirements 24
• 2.4.3 Model upgrade to TensorFlow 2 and performance degradation analysis 25
• 2.4.4 Dataset refinement and reorganization of pest and disease class taxonomy 26
• 2.4.5 Class balancing through large-scale image augmentation 26
• 2.4.6 Final model training using the refined dataset (V2-Refined) 27
• 2.4.7 Integrated diagnostic engine and extensibility 28
• 2. 5. Deep learningbased automatic beating cycle estimation from video data 29
• 2.5.1 Video preprocessing and brightness-based signal extraction 29
• 2.5.2 Deep learning architecture for spatiotemporal pattern learning 30
• 2.5.3 Model training and implementation details 30
• 2.5.4 Performance evaluation metrics 31
• 2.5.5 Validation through temporal signal consistency 31
• 2.5.6 System scalability and real-time applicability 32
• 2. 6. Deep learningbased sperm detection and tracking from video data 33
• 2.6.1 Overview of the CASA video analysis pipeline 33
• 2.6.2 Video acquisition and preprocessing 33
• 2.6.3 Deep learningbased sperm detection model 34
• 2.6.4 Multi-object tracking using DeepSORT 34
• 2.6.5 Quantitative evaluation of detection and tracking performance 35
• 2.6.6 Trajectory reconstruction and WHO motility parameter computation 35
• 2.6.7 Computational performance and real-time processing capability 36
• 2.6.8 Implementation details and system scalability 36
• 3. Results and Discussion 37
• 3.1 Characteristics of structured data and analytical methodology 37
• 3.2 Machine learning prediction based on experimental data 41
• 3.2.1 Model performance and predictive behavior 43
• 3.2.2 Identification of optimal process conditions 44
• 3.2.3 Experimental validation and physical interpretation 46
• 3.2.4 Implications, limitations, and future directions 48
• 3.3 Machine learning prediction using open-source data 49
• 3.3.1 Model performance across structurally distinct datasets 52
• 3.3.2 Sequential performance prediction and interpretation 54
• 3.3.3 Comparison between model predictions and experimental observations 56
• 3.3.4 Implications, limitations, and outlook 58
• 3.4. Machine learning prediction using open data and generative AI 60
• 3.4.1 Structural limitations of conventional human-coded machine learning 60
• 3.4.2 Conceptual innovation of generative AIbased no-code machine learning 62
• 3.4.3 Quantitative comparison between human-coded and no-code machine learning 64
• 3.4.4 Research significance and emergence of a new analytical paradigm 66
• 3.5 Deep Learning Analysis and Applications Based on Image and Video Data 68
• 3.6 Image-Based Pest and Disease Diagnosis Using Deep Learning 71
• 3.6.1 Initial Dataset Construction and Baseline Model Performance 71
• 3.6.2 Impact of Data Imbalance and Structural Bias on Model Generalization 73
• 3.6.3 Dataset Expansion and Unexpected Performance Degradation 74
• 3.6.4 Data Refinement and Class System Reorganization 76
• 3.6.5 Class Balancing Through Large-Scale Image Augmentation 77
• 3.6.6 Performance Recovery and Improvement with the Refined Dataset 77
• 3.6.7 Dominance of Data Quality Over Model Complexity 79
• 3.6.8 Implications and Future Directions 79
• 3.7 Deep LearningBased Automatic Beat Cycle Estimation from Video Data 80
• 3.7.1 Automated Beat Signal Extraction from Cardiomyocyte Videos 80
• 3.7.2 CNNLSTM Hybrid Model Performance and Temporal Pattern Learning 82
• 3.7.3 Quantitative Performance Analysis and Computational Efficiency 82
• 3.7.4 Implications for High-Throughput Cardiac Screening 84
• 3.7.5 Conclusive Significance and Future Applicability 85
• 3.8 Deep LearningBased Object Detection and Tracking Using Video Data 86
• 3.8.1 Automated Analysis of Sperm Motility from Microscopy Videos 86
• 3.8.2 Performance of Deep LearningBased Object Detection and Tracking 88
• 3.8.3 Quantitative Analysis of Sperm Motility Parameters 90
• 3.8.4 Processing Speed and Real-Time Analysis Capability 91
• 3.8.5 Clinical Significance and Implications for Digital Reproductive Analysis 92
• 3.8.6 Summary and Future Outlook 96
• 4. Conclusion 97
• References 101
• 초록 105