Title Page

Contents

ABSTRACT 9

Ⅰ. Introduction 11

Ⅱ. Patients and Methods 14

1. Study Population and Design 14

2. Information on Collected Data 15

1) Clinical information 15

2) FDG PET/CT image acquisition 15

3. PET Image Analysis 16

1) Tumor segmentation 16

2) Feature extraction 17

3) Data transformation 17

4. Predictive Model Construction 18

1) FDG radiomic feature selection and RAD-score construction. 18

2) Independent prognostic value of RAD-score 18

3) Model construction and predictive performance 19

5. Statistical Analysis Methods 19

Ⅲ. Results 20

1. Baseline Characteristics of Patients 20

2. Construction of Radiomic Model 21

1) Feature selection and formulating RAD-scores 21

2) Independent prognostic value of constructed RAD-scores 22

3. Construction of Clinical and Combined Model 22

4. Evaluation of Predictive Performance of Constructed Models 23

1) Predictive performance of clinical and radiomic model 23

2) Predictive performance of combined model 24

5. Model Evaluation in the Validation Set 24

1) The impact of data harmonization 24

2) Predictive performance of constructed models in the validation set 25

6. Representative Clinical Case Examples 26

Ⅳ. Discussion 26

1. Developing a Radiomic Model for NSCLC Survival Prediction 26

2. Comparative Studies on FDG PET-based Radiomics 27

3. Methodological Considerations in FDG PET/CT Imaging Analysis 29

4. Significant PET Parameters for Prediction of Survival 29

1) Significance of Conventional PET Parameters 29

2) Significance of Texture PET Parameters 30

5. Impact of Tumor Segmentation Method 31

6. Challenges in High-Dimensional Data Analysis 31

7. Validation and Harmonization in Predictive Modeling 32

8. Performance Discrepancies in Training and Validation Cohorts 33

9. Integration of Clinical Variables in Predictive Modeling 33

10. Clinical implication of our Predictive Model 34

11. Limitations and Future Directions 35

Ⅴ. Conclusion 36

References 37

국문초록 59

Table 1. Clinical characteristics of the training cohort and the validation cohorts. 45

Table 2. Cox regression analysis of PET parameters for overall survival in the training cohort. 46

Table 3. Cox regression analysis of clinical variables for overall survival in the training cohort. 46

Table 4. Multivariable Cox regression analysis of combined model for overall survival in the training cohort. 47

Table 5. Comparative time-dependent AUROC performance of clinical, radiomic, and combined model in the training cohort. 47

Table 6. Comparison of performance of radiomic scores based on the application of harmonization method in the validation cohort. 48

Table 7. Univariable Cox regression analysis for overall survival in the validation cohort. 48

Table 8. Comparative time-dependent AUROC performance of clinical, radiomic, and combined model in the validation cohort. 49

Supplementary Table 1. List of 72 quantitative PET-based radiomic features. 58

Figure 1. Work flow of the present study. 50

Figure 2. Examples of two tumor segmentation methods. 51

Figure 3. Demographic and radiomic feature selection using the LASSO Cox method for gradient-based segmentation (A) and fixed threshold-based segmentation (B). 52

Figure 4. Kaplan-Meier survival curves comparing OS between two groups stratified by RAD-score 2.5 (A), RAD-score Edge (B), age (C), and tumor stage (D). 53

Figure 5. Time-dependent ROC curves comparing the predictive capacity of the three models. 54

Figure 6. Comparisons of RAD-score Edge (A) and RAD-score 2.5 (B) between the training set and the validation set after the application of harmonization method. 55

Figure 7. A 58-year-old male NSCLC patient with relatively high RAD-scores. 56

Figure 8. A 66-year-old male NSCLC patient with relatively low RAD-scores. 57