Feature Ranking Importance from Multimodal Radiomic Texture Features using Machine Learning Paradigm: A Biomarker to Predict the Lung Cancer

The machine learning based techniques for detection of lungs cancer can assist the clinicians in assessing the risk of pulmonary nodules being malignant. We are developing non-invasive methods to accurately distinguish the non-small cell cancer carcinoma (NSCLC) from small cell cancer carcinoma (SCLC) brain metastases. In this study, we extracted multimodal radiomic features including texture and statistical Haralick texture, gray level co-occurrence matrix (GLCM) features, Gray level size-zone matrix (GLSZM) features, Gray-level run-length matrix (GLRLM) features. We also applied image enhancement contrast stretching and gamma correction to further improve the classification performance. We then ranked these features in order to investigate that which features category is more important to accurately distinguish the lung cancer subtypes. We employed robust machine learning techniques. We evaluated the performance based on top ranked 03 and 05 features and last ranked 05 and 02 features based on the receiver operating curve (ROC). The highest classification performance in terms of accuracy and AUC was obtained with all Haralick texture features using SVM polynomial with accuracy (99.89%) and AUC (0.9984). The classification performance with contrast stretching [0.02, 0.08; 0.05, 0.95] and gamma correction with gamma = 0.5 yielded highest accuracy of 100% and AUC of 1.00. The top three ranked features using image enhancement methods also yielded accuracy more than 95% which indicates that these top ranked features contributed higher in accuracy classifying the lung cancer subtypes. The results revealed that proposed model with multimodal features, image enhancement techniques and features ranking methods improved the classification performance which can be used for better diagnostic aid to improve the decision making to treat the patients suffering from SCLC and NSCLC.