A Study on Early Detection Methods of Lung Cancer Integrated with Diffusion Models = 확산 모델을 접목한 조기 폐암 검출 방법에 관한 연구

다국어 초록 (Multilingual Abstract)

Lung cancer is a malignant disease that poses a serious threat to human health. As a key indicator of early-stage lung cancer, the precise identification of pulmonary nodules is crucial for improving patient survival rates. Computed tomography is currently the primary method for early lung cancer screening. In the early detection, accurate classification of pulmonary nodules relies on precise lesion localization or effective segmentation , which in turn depends on high-quality images. Denoising, segmentation, and classification of computed tomography images are core components in the construction of intelligent diagnostic assistance systems. Leveraging the powerful nonlinear modeling capabilities of deep neural networks to automatically extract high-level semantic features, deep learning techniques have demonstrated significant advantages in computed tomography image denoising, segmentation, and classification tasks. In recent years, diffusion models have exhibited high fidelity and detail recovery capabilities, offering new technical pathways for optimizing computed tomography image quality and enhancing lung nodule segmentation accuracy.
This paper focuses on pulmonary computed tomography images, conducting systematic research across three critical medical image processing domains: Low-dose computed tomography image denoising, pulmonary nodule segmentation, and pulmonary nodule classification. The objective is to design high-performance methods that address the challenges posed by the complex and variable manifestations of pulmonary nodules in terms of morphology, texture, location, and boundaries. The study thoroughly analyzes the characteristics of lung image processing across different tasks. Based on the specific requirements for extracting key feature information for each task, appropriate deep learning models and methods are selected to achieve the corresponding objectives. The effectiveness of the introduced methods is validated through experiments. The main research content of this paper encompasses three aspects:
(1) To address issues such as poor image quality, blurred structural details, and texture loss in lung Low-dose computed tomography images, a denoising method based on a mean-reversion-driven diffusion model was put forward. This approach replaces the standard stochastic differential equation with a mean-reversion stochastic differential equation in the forward process, stabilizing the noise diffusion direction. In the backward process, Low-dose computed tomography images are concatenated with pure noise images before feeding into the decoder of the denoising network. This helps the network learn the differences between image noise patterns and tissue textures. An adaptive weight attention mechanism is introduced in the skip connection to enhance the decoder's ability to recover structural details, thereby reconstructing high-quality images matching the input size. Experimental results on the public Mayo grand challenge 2016 dataset demonstrate that this method effectively restores computed tomography image clarity with good fidelity performance.
(2) Addressing the challenges of small lung nodules with low contrast against surrounding tissues, we propose a lung nodule segmentation method integrating a conditional diffusion model. This method, based on Denoising Diffusion Probabilistic Model, improves the denoising network of the diffusion model. It designs a lightweight conditional feature module that encodes the original computed tomography images into conditional features, which are then dynamically fused with the input features through a designed gated fusion module. To better restore edge details in segmentation masks, attention-enhancing modules are introduced at skip connections between encoder and decoder layers to suppress extraneous noise. Additionally, a joint loss function is designed to improve segmentation performance for blurred boundaries and minute nodules in complex scenarios. Experiments on the LUNA16 public dataset demonstrate that this enhanced segmentation method achieves high accuracy and exhibits outstanding performance.
(3) To enhance feature representation of pulmonary nodules given their morphological, size, and textural variations in computed tomography images, we propose a nodule classification method based on a global multi-scale fusion attention mechanism. The backbone network comprises multiple multi-scale modules, each containing multi-scale convolutional layers. After feature extraction at different scales, max pooling is applied. Features extracted from each module are further fed into a global feature fusion module to obtain high-level semantic features across different network depths. These features are then enhanced via the attention mechanism before being input to a fully connected layer. Finally, a classification layer outputs the probabilities of benign and malignant lung nodules. Experimental results on the LIDC-IDRI public dataset demonstrate that the introduced global multi-scale classification network achieves the best performance in both accuracy and sensitivity metrics, validating its effectiveness and applicability for lung nodule classification tasks.

목차 (Table of Contents)

• ABSTRACT VIII
• Chapter 1 Introduction 1
• 1.1 Research Background and Significance 1
• 1.2 Current Status of Domestic and International Research 4
• 1.2.1 Research Status of LDCT Image Denoising 7
• 1.2.2 Research Status of Lung Nodule Segmentation 11
• 1.2.3 Research Status of Lung Nodule Malignancy Classification 16
• 1.2.4 Application Status of Diffusion Models in Medical Imaging 20
• 1.3 Research Content and Objectives 22
• 1.4 Thesis Organization 25
• Chapter 2 Related Theories and Foundations 28
• 2.1 An Overview of Medical Imaging 28
• 2.1.1 CT images 30
• 2.1.2 Window width and window level 32
• 2.1.3 Flow chart of CT image processing 33
• 2.2 Theoretical basis of deep learning 34
• 2.2.1 Artificial Neural Networks 35
• 2.2.2 Convolutional Neural networks 36
• 2.2.3 U-Net network 41
• 2.2.4 Attention mechanism 44
• 2.3 Diffusion models 46
• 2.3.1 Denoising Diffusion Probabilistic Model DDPM 47
• 2.3.2 Conditional denoising diffusion probability model 50
• 2.4 Introduction to the Lung CT Image Dataset 51
• 2.4.1 LIDC-IDRI 52
• 2.4.2 LUNA16 53
• 2.4.3 Mayo Grand Challenge 54
• 2.5 Summary of This Chapter 54
• Chapter 3 Denoising Method for Pulmonary LDCT Images Based on Mean-Reverting Driven Diffusion Model 56
• 3.1 Introduction 57
• 3.2 LDCT Image Denoising Network Model 60
• 3.2.1 Overall Network structure 60
• 3.2.2 Mean-Reverting Stochastic Differential Equations 61
• 3.2.3 Adaptive Weighted Fusion Spatial-Channel Attention Module 64
• 3.3 Loss Function and Evaluation Metrics 68
• 3.3.1 Loss Function 68
• 3.3.2 Evaluation Metrics 69
• 3.4 Experiments and Analysis 71
• 3.4.1 Data Preprocessing and Experimental Setup 71
• 3.4.2 Experimental Results and Comparative Analysis 73
• 3.4.3 Ablation Experiments 77
• 3.5 Summary of this chapter 79
• Chapter 4 Lung Nodule Segmentation Method Integrating Conditional Diffusion Model 81
• 4.1 Introduction 83
• 4.2 Segmentation Network Integrating Conditional Diffusion Model 86
• 4.2.1 Overall Network Structure 86
• 4.2.2 Gated Fusion Module 88
• 4.2.3 Attention Enhanced Module 89
• 4.3 Loss Function and Evaluation Metrics 91
• 4.3.1 Loss Function 91
• 4.3.2 Evaluation Metrics 92
• 4.4 Experiments and Analysis 95
• 4.4.1 Data Preprocessing and Experimental Setup 95
• 4.4.2 Experimental Results and Comparative Analysis 96
• 4.4.3 Ablation Studies 100
• 4.5 Summary of This Chapter 103
• Chapter 5 Pulmonary Nodule Classification Method Based on Global Multi-scale Fusion Attention Mechanism 105
• 5.1 Introduction 106
• 5.2 Pulmonary Nodule Classification Network Structure 109
• 5.2.1 Overall Network Structure 109
• 5.2.2 Multi-scale Layer 110
• 5.2.3 Global Feature Fusion Module 112
• 5.2.4 Simplified Attention Module 113
• 5.3 Loss Function and Evaluation Metrics 115
• 5.3.1 Loss Function 115
• 5.3.2 Classification Evaluation Metrics 117
• 5.4 Experiments and Analysis 118
• 5.4.1 Data Preprocessing and Experimental Setup 118
• 5.4.2 Experimental Results and Comparative Analysis 119
• 5.4.3 Ablation Study 123
• 5.5 Summary of This Chapter 124
• Chapter 6 Summary and Outlook 126
• 6.1 Summary of This Work 126
• 6.2 Future Work and Outlook 128
• REFERENCE 131