Predicting Ki-67 expression levels in non-small cell lung cancer using an explainable CT-based deep learning...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Ki-67 expression is a crucial biomarker for non-small cell lung cancer (NSCLC) proliferation and prognosis.
• Accurate prediction of Ki-67 levels is essential for guiding personalized therapeutic strategies in NSCLC patients.

Purpose of the Study:

• To develop and validate an interpretable model for predicting Ki-67 expression levels in NSCLC.
• To combine clinical-radiological, radiomic, and deep learning features for enhanced predictive performance.

Main Methods:

• Retrospective study involving 259 NSCLC patients (training/validation) and 112 (independent test set).
• Extraction of radiomic and deep learning features (ResNet18) from CT images.
• Development of four support vector machine models: clinical-radiological, radiomic, deep learning, and a combined model, utilizing LASSO for feature selection and SHAP for interpretability.

Main Results:

• The combined model demonstrated superior predictive performance in the independent test set with an AUC of 0.892.
• The combined model significantly outperformed individual clinical-radiological (AUC 0.820), radiomic (AUC 0.750), and deep learning (AUC 0.817) models.
• SHAP analysis highlighted deep-score, histological type, and rad-score as key predictors of Ki-67 expression.

Conclusions:

• An interpretable combined model effectively predicts Ki-67 expression in NSCLC.
• This imaging-based approach offers valuable evidence for optimizing personalized treatment strategies in NSCLC.