ViT-CNN Fusion with Attention-Based Ensemble for Lung Cancer Classification in Low-Resource Settings

다국어 초록 (Multilingual Abstract)

Background: Lung cancer remains one of the most lethal malignancies worldwide, where delayed diagnosis is closely associated with poor prognosis. Accurate and timely classification of lung cancer subtypes is critical for guiding appropriate treatment strategies and improving survival outcomes. Methods: This thesis proposes a hybrid deep learning framework that fuses Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) using an attention-based ensemble mechanism for multi-class lung cancer classification from chest CT images in low-resource settings. CNNs are employed to capture discriminative local spatial features, while ViTs model long-range global dependencies. The model is trained and evaluated on 3,690 de-identified CT images from APEX Clinic, Uzbekistan. Results: The proposed hybrid CNN–ViT ensemble achieves a classification accuracy of 99.19% and a macro-AUC of 0.9997, demonstrating superior or comparable performance to standalone baselines. Conclusion: These findings indicate that synergistic integration of CNN and ViT architecture provides a reliable framework for computer-aided lung cancer diagnosis in resource-constrained clinical environments.

목차 (Table of Contents)

• CHAPTER 1. Introduction 10
• 1.1 Background 10
• 1.2 Problem statement 10
• 1.3 Research motivation 11
• 1.4 Research objectives 11
• 1.5 Scope of the study 11
• 1.6 Thesis structure 12
• CHAPTER 2. Literature view 14
• 2.1 Introduction 14
• 2.2 Convolutional Neural Networks (CNNs) in medical imaging 14
• 2.3 Vision Transformers (ViTs) in medical imaging 15
• 2.4 Hybrid CNN–ViT architectures 16
• CHAPTER 3. Dataset and preprocessing 18
• 3.1 Data set 18
• 3.2 Ethical considerations 19
• 3.3 Data processing 19
• 3.4 Data labelling 19
• CHAPTER 4. Proposed methodology 22
• 4.1 Overview of the proposed architecture 22
• 4.2 CNN Feature extraction 22
• 4.3 Vision transformer feature extraction 23
• 4.4 Attention-based feature fusion module 23
• 4.5 Classification head 23
• 4.6 Model optimization for low-resource environments 24
• 4.7 Advantages over conventional approaches 24
• 4.8 Implementation environment and software configuration 24
• 4.8.1 Software environment 24
• 4.8.2 Hardware environment 25
• 4.8.3 Execution environment and reproducibility settings 25
• 4.8.4 Deployment and considerations 26
• CHAPTER 5. Experiments and results 27
• 5.1 Model performance metrics 27
• 5.2 Confusion matrix analysis 28
• 5.3 ROC-AUC analysis 30
• 5.4 Comparative evaluation 33
• CHAPTER 6. Discussions 34
• 6.1 Overview of experimental findings 34
• 6.2 Comparative analysis of model architectures 35
• 6.2.1 Convolutional Neural Network (CNN) 35
• 6.2.2 Vision Transformer (ViT) 35
• 6.2.3 Hybrid CNN–ViT ensemble 36
• 6.3 Clinical significance and practical implications 36
• 6.4 Limitations 37
• 6.5 Future research directions 38
• 6.6 Summary 39
• CHAPTER 7. Conclusion 40
• References 41
• 국문초록 44