# Predicting EGFR gene mutation in lung adenocarcinoma using spectral CT combined with AI parameters: a diagnostic accuracy study

## TL;DR

This study shows that combining spectral CT and AI parameters with clinical data can noninvasively predict EGFR mutations in lung cancer patients, potentially reducing the need for biopsies.

## Contribution

The novel contribution is a predictive model integrating spectral CT, AI-derived features, and clinical indicators for noninvasive EGFR mutation detection in lung adenocarcinoma.

## Key findings

- Smoking history, λHU, and tumor surface area were identified as independent predictors of EGFR mutation.
- The predictive model achieved a moderate diagnostic accuracy with an AUC of 0.713.
- The model showed good calibration and clinical utility for noninvasive EGFR mutation prediction.

## Abstract

Epidermal growth factor receptor(EGFR) mutation is one of the most critical biomarkers in non-small cell lung cancer (NSCLC), holding significant clinical implications for guiding targeted therapy selection and prognostic assessment in patients. This study aims to evaluate the predictive value of spectral CT parameters, artificial intelligence (AI)-derived parameters, and clinical indicators for EGFR mutation in lung adenocarcinoma.

This retrospective study analyzed 150 patients with pathologically confirmed lung adenocarcinoma. All patients underwent EGFR genotyping, non-contrast CT, and spectral contrast-enhanced CT. Spectral parameters included spectral curve slope (λHU), iodine concentration (IC), water concentration (WC), Effective atomic number (Effective-Z), and CT values at 70 keV. An AI-assisted diagnostic system automatically extracted quantitative AI parameters: The three-dimensional (3D) radiomic features(including long-axis diameter, short-axis diameter, surface area, 3D long-axis diameter, maximum cross-sectional area, volume), CT attenuation histogram features(including solid component percentage, mean CT value, median CT value, CT value standard deviation, maximum CT values, minimum CT values, kurtosis, skewness, energy, and entropy)and morphological characteristics(including compactness, sphericity). Correlations between spectral CT parameters, AI parameters, clinical variables, and EGFR mutation status were assessed. Independent predictors were identified via multivariate analysis to construct a predictive model.

Univariate analysis revealed associations between EGFR mutation and gender (P = 0.013), smoking history (P = 0.001), λHU (P = 0.049), and tumor surface area (P = 0.043). Multivariate analysis identified smoking history (P = 0.012), λHU (P = 0.015), and surface area (P = 0.029) as independent predictors. The predictive model integrating these three factors achieved an AUC of 0.713 (95% CI: 0.628–0.797), a specificity of 0.754, and a sensitivity of 0.600, demonstrating moderate diagnostic accuracy. Calibration curves indicated good agreement between predicted and observed probabilities, while decision curve analysis confirmed clinical utility.

The integration of spectral CT and AI-derived quantitative parameters with clinical indicators demonstrates significant potential for noninvasive prediction of EGFR mutation in lung adenocarcinoma. This non-invasive predictive approach could reduce unnecessary invasive biopsies. Particularly in patients with contraindications to invasive procedures, this model offers a viable alternative.