Non-Small Cell Lung Cancer Recurrence-Risk Prediction with Radiogenomics Features Using Transfer Learning

Non-small cell lung cancer (NSCLC) is the most prevalent subtype of lung cancer. Cancer recurrence is the event of re-emergence of tumor after surgery, which occurs in more than 30% of NSCLC cases and reduces patient survivability. The prediction of risk of postoperative recurrence in patients at preoperative stages can be beneficial for doctors. In this work, the aim was to integrate preoperative radiological and genomics data and apply the transfer learning strategy to develop machine learning (ML) models for the prediction of postoper-ative recurrence-risk in NSCLC patients. The chest computed tomography (CT) scan images and gene expression data of 117 patients were obtained from a publicly available dataset, namely NSCLC-Radiogenomics. Two pre-trained convolutional neural networks, Inception and Xception, were used to extract features from the CT scan images. ML models for recurrence-risk prediction were developed with a combination of the extracted radiological features and the genomic features using support vector machine (SVM), random forest (RF), and multi-layer perceptron (MLP) algorithms. The SVM model trained with Xception-extracted radiological features and genomic features achieved an accuracy of 99% in 5-fold cross validation, and the optimal external validation performance with 97% accuracy. This optimal model developed with the radiogenomics features achieved a performance gain than the models developed with only radiological or only genomics (unimodal) features. In summary, ML models were developed with the transfer learning strategy using radiogenomics features for the prediction of postoperative recurrence-risk before surgery to aid the clinicians in managing the postoperative therapy of NSCLC patients.