ESTRO 2024 - Abstract Book

S4994

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Results:

The discrimination performance of Model-L (C-index: 0.71, 0.62, 0.67 for training set, test-set-1, and test-set-2) was comparable to Model-T (C-index: 0.68, 0.63, 0.65). Notably, Model-CLT (C-index: 0.82 for the training set, 0.76 for test set-2) surpassed Model-LT (C-index: 0.78, 0.73, 0.73). XGBoost models consistently outperformed Cox models. Risk stratification based on the developed models revealed distinct prognosis for different risk groups (log-rank, P<0.05). Dependence plots effectively illustrated the relationships between radiomics/ clinical parameters, and OS.

Training set

Test-set-1

Test-set-2

C-index

95% CI

C-index

95% CI

C-index

95% CI 0.52 0.73 0.51 0.74 0.64 0.82

Conclusion:

Model

Model-L

0.71

0.64-0.75

0.62

0.55-0.67

0.67

The

combination

of

radiomics clinical features in AI model demonstrated effective and

Model-T

0.68

0.65-0.73

0.63

0.56-0.69

0.65

Model-LT

0.78

0.72-0.82

0.73

0.65-0.79

0.73

stratification of patients into different risk groups. Both lung and tumor-based radiomics features exhibited significant prognostic value, highlighting the potential of AI-driven approaches in enhancing personalized treatment strategies. Model CLT 0.82 0.75-0.88 - - 0.76 0.64 0.87

Keywords: lung cancer, explainable AI, prognosis