Machine learning–based 18F-FDG PET/MR fusion radiomics for predicting EGFR mutation in lung cancer

<title>Abstract</title> <p> Background This study aimed to develop a machine learning–based radiomics model based on 18F‑FDG PET/MR images, to predict EGFR mutation status and mutation abundance in lung cancer, and to explore the value of multimodal radiomic features in clinical decision‑making. Methods 66 patients were included in this retrospective study. Then, the registered PET and MR images were fused using 3D Slicer. Radiomic features were extracted from the PET/MR images and fused images using PyRadiomics. Feature selection was conducted using LASSO regression with leave-one-out cross-validation (LOOCV). Logistic regression was employed for binary classification of EGFR mutation status, and linear regression was used to predict continuous variant allele frequency. Results EGFR mutation status was significantly associated with female sex, higher T stage and advanced pathological stage, irregular nodule shape, larger tumor dimensions, and greater MR signal heterogeneity (all <italic>p</italic>  &lt; 0.05). For predicting EGFR mutation abundance, the radiomics model performed best in the overall cohort (R² = 0.687), with PET standardized metabolic activity ( <italic>β</italic>  = 18.46, <italic>p</italic>  = 0.011) and PET wavelet LLH kurtosis ( <italic>β</italic>  = 7.23, <italic>p</italic>  = 0.009) emerging as key predictors. In the Exon21 subgroup, a single feature, PET wavelet‑HHL ngtdm Busyness, explained 85.8% of the variance in variant allele frequency (R² = 0.858). </p>