Purpose Gene mutations are mutually exclusive in non-small cell lung cancer (NSCLC). Using EGFR and KRAS as examples, this study aims to assess the significance of mutation mutual exclusion information in the optimization of radiomics algorithms for predicting gene mutation.Methods We retrospectively analyzed 161 NSCLC patients with 18F-FDG PET/CT scans and EGFR and KRAS gene mutation data. Patients were randomly divided into training and testing cohorts. The Pyradiomics toolkit was used for radiomics feature extraction. Based on these features, radiomics score (RS) models were developed for predicting KRAS mutations using the gradient boosting decision tree (GBDT) algorithm. Furthermore, to investigate the value of adding mutation mutual exclusion information, a composite model combining PET/CT RS and EGFR mutation status was developed using logistic regression. The area under the curve (AUC), specificity, sensitivity, and accuracy were calculated for model performance evaluation in the training and test cohorts. To test the generalizability of this optimization method, models for predicting EGFR mutation were established in parallel, with or without adding KRAS gene mutation information.Results Compared with CT, the PET/CT based RS model exhibited higher AUC (KRAS: 0.792 vs 0.426; EGFR: 0.786 vs 0.644). By integrating EGFR mutation information into the PET/CT RS model, the AUC, accuracy, and specificity for predicting KRAS mutations were all elevated in the test cohort (0.928, 0.857, 0.897 vs 0.792, 0.755, 0.769). Conversely, the composite model for predicting EGFR mutations could also be optimized by adding KRAS gene mutation information (AUC, accuracy, and specificity: 0.877, 0.776, 0.700 vs 0.786, 0.694, 0.567). By adding EGFR and KRAS exclusive mutation information, respectively, the composite model corrected 55.4% and 30.7% false positive cases produced by the PET/CT RS model in the test cohort, without sacrificing sensitivity.Conclusion Integrating the mutation status of a known gene is a potential method to optimize radiomics models for predicting another gene mutation. This method may help predict unconventional gene mutations when the second biopsy is clinically difficult to carry out.