High-fidelity super-resolution CT radiomics for non-invasive EGFR mutation prediction in lung adenocarcinoma: a multi-center pooled analysis

To develop and validate a high-fidelity super-resolution (SR)-enhanced radiomics framework using a Residual Channel Attention Network (RCAN) for non-invasive EGFR mutation prediction in lung adenocarcinoma (LUAD).

This retrospective multi-center study included 373 patients, partitioned into training (n = 298) and testing (n = 75) sets. CT images were reconstructed to a 1024 × 1024 matrix via RCAN to restore latent high-frequency textures. A standardized pipeline-including ComBat harmonization, ICC-based fidelity filtering, and LASSO regression-was employed to extract and select resolution-invariant features. Five machine learning classifiers were evaluated, and a combined nomogram integrated the SR-enhanced signature with clinical predictors. Model performance was assessed using AUC, DeLong tests, and decision curve analysis (DCA), with interpretability provided by SHAP analysis.

The SR-enhanced model significantly outperformed the original-resolution (OR) baseline, increasing the AUC from 0.60 (95% CI: 0.47-0.74) to 0.84 (95% CI: 0.75-0.93) in the testing set (P < 0.001). Consistent performance was maintained across imaging centers (P = 0.555) and histological subtypes. The combined nomogram achieved a robust AUC of 0.86 (95% CI: 0.78-0.94), demonstrating superior calibration and clinical net benefit. SHAP analysis revealed that glszm_ZoneVariance-a marker of intratumoral heterogeneity-was the predominant predictor revealed via SR reconstruction.

RCAN-driven SR reconstruction effectively addresses CT resolution limitations, capturing fine-grained radiogenomic signatures critical for molecular phenotyping. This high-fidelity framework offers a robust, non-invasive decision-support tool for personalized precision oncology in LUAD.