A Radıogenomıcs-Based Approach to Clınıcal Decısıon Making
Özet
RadioGenomics is an emerging research field focusing on the integrated analysis of medical images and genetic information. This thesis aims to evaluate the efficacy of image related features (radiomics) and genetic features (genomics) in classifying the histology type (adenocarcinoma or squamous cell carcinoma) of non-small cell lung cancer patients. Computed tomography images and RNA sequencing gene expression data from the same patients were obtained from a publicly available dataset. A simulation study was conducted based on the properties of this real dataset. Two scenarios were included to assess the effect of class distribution (unbalanced and balanced). Analyses were conducted to examine the impact of feature selection on classification performance, comparing results with and without performing feature selection methods. Elastic Net, Random Forest, XGBoost and Support Vector Machines with three different kernel functions were used as classification methods. The means of model performances on test set were compared under different scenarios. The findings of the simulation study indicate that unbalanced scenarios exhibited higher sensitivity and F1 scores than balanced scenarios, as algorithms tend to predict the majority class more frequently in unbalanced scenarios. There were no substantial differences between the scenarios with and without feature selection. Considering the characteristics of the dataset used in the simulation study, the diagnostic ability could potentially be improved by combining genetic data in addition to image data. However, when genomics dataset is already available, the integration of radiomics data may not be needed.
