Prediction of ALK, EGFR and KRAS mutations in lung adenocarcinoma based on PET/CT radiomics by cluster and synthetic minority over-sampling technique

Radiomics holds great potential for the noninvasive evaluation of lung cancer mutations; however, challenges related to data imbalance and model robustness limit its current efficacy and safety. This multicenter study aimed to develop and validate a model utilizing the Synthetic Minority Over-sampling Technique (SMOTE) and clustering methods to address the challenges associated with small sample sizes and imbalanced radiomics features between negative and positive samples. A total of 230 patients with lung adenocarcinoma who underwent positron emission tomography/computed tomography and genetic testing for anaplastic lymphoma kinase (ALK), epidermal growth factor receptor (EGFR), or Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations were included. Multi-layer perceptron (MLP) models were developed for each genetic mutation group based on 2 different approaches for processing radiomics features. The MLP model incorporating SMOTE and clustering demonstrated potentially improved performance, with AUC values of 0.591 for ALK, 0.750 for EGFR, and 0.789 for KRAS in the external test set. These findings suggest that the MLP model combined with SMOTE and clustering has potential for predicting ALK, EGFR, and KRAS mutations in lung adenocarcinoma.