Radiomics-machine learning model for predicting invasiveness of subcentimeter subsolid lung adenocarcinoma: a...

Area of Science:

• Pulmonary Medicine
• Radiology
• Artificial Intelligence

Background:

• Distinguishing invasive adenocarcinoma from pre-invasive lesions in subcentimeter subsolid nodules (SSNs) using computed tomography (CT) is challenging.
• An interpretable radiomics-machine learning (ML) model using SHapley Additive exPlanations (SHAP) was developed to predict invasiveness.

Purpose of the Study:

• To develop and validate an interpretable radiomics-ML model for predicting invasiveness in subcentimeter SSNs.
• To leverage SHAP for model interpretability and identification of key predictive features.

Main Methods:

• A two-center retrospective study included 177 patients (training/internal validation) and 83 (external validation) with surgically confirmed lung adenocarcinoma SSNs (≤1 cm).
• Radiomic features were extracted, and mRMR and LASSO regression selected predictive features.
• Logistic regression (LR), random forest (RF), and support vector machine (SVM) models were trained and validated, with the best model assessed externally.
• SHAP analysis provided global and local model interpretability.

Main Results:

• Ten radiomic features were selected. The LR model achieved optimal internal validation performance (AUC: 0.842) and demonstrated robust external validation (AUC: 0.778).
• The LR model outperformed RF and SVM, showing superior generalizability and clinical utility confirmed by decision curve analysis (DCA).
• SHAP identified key predictors: wavelet_HLL_glszm_LowGrayLevelZoneEmphasis, original_shape_Flatness, and log_firstorder_LoG.Minimum.

Conclusions:

• The developed radiomics-LR model accurately predicts invasiveness in subcentimeter SSNs.
• The model's interpretability via SHAP provides biologically plausible explanations, supporting clinical integration for surgical decision-making.