Advancing NSCLC pathological subtype prediction with interpretable machine learning: a comprehensive radiomics-based...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Non-small cell lung cancer (NSCLC) has multiple histological subtypes requiring accurate diagnosis.
• Computed Tomography (CT) imaging is crucial for lung cancer assessment.
• Interpretable machine learning offers potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate interpretable machine learning models for diagnosing NSCLC histological subtypes.
• To utilize radiomic features extracted from CT images for subtype classification.
• To enhance clinical decision-making through accurate and interpretable AI models.

Main Methods:

• A retrospective cohort of 317 NSCLC patients was used, with data split into training (222) and validation (95) sets.
• 1,834 radiomic features were extracted, with feature selection using statistical methods and Synthetic Minority Over-sampling Technique (SMOTE) for data imbalance.
• Six classifiers (Logistic Regression, SVM, Decision Tree, Random Forest, XGBoost, LightGBM) were trained and evaluated using accuracy and AUC metrics, with SHAP for interpretability.

Main Results:

• Key radiomic features (9 for ADC, 12 for SCC, 8 for LCC) were identified for subtype prediction.
• The XGBoost model achieved AUCs of 0.789 for SCC and 0.848 for LCC.
• The Random Forest model achieved an AUC of 0.748 for adenocarcinoma (ADC) prediction.

Conclusions:

• Machine learning models utilizing CT imaging demonstrate strong predictive performance for NSCLC subtypes (SCC, LCC, ADC).
• The interpretability of these models supports their integration into clinical decision-making workflows.
• This approach offers a valuable tool for precise histological subtype diagnosis in NSCLC.