Evaluation of Radiomics and Machine Learning for Classifying Pulmonary Nodules in CT Images
Abstract
Purpose: Lung cancer is a deadly disease that has high occurrence and death rates, worldwide. Clinicians are widely using computed tomography imaging for the detection of lung cancer. Radiomics extracted from medical images together with a machine learning platform has given encouraging results in lung cancer diagnosis. Therefore, this study is proposed with the aim of efficiently applying and evaluating radiomics and ML techniques to classify pulmonary nodules in CT images.
Materials and Methods: Lung Image Data Consortium is utilized in which nodules are given malignancy scores 1 through 5 i.e. benign through malignant. Three scenarios are created using these scores: G54 Vs G12, G543 Vs G12, and G54 Vs G123. Radiomics is extracted using Shape, Gray Level Co-occurrence Method, Gray Level Difference Method, and Gray Level Run Length Matrix along with Wavelet Packet Transform. To select a relevant set of features, four techniques i.e. Chi-square test, Analysis of variance, boosted ensemble classification tree and bagged ensemble classification tree are applied. The classification of nodules into benign or malignant is evaluated by using six models of support vector machine.
Results: The results, in Scenario 1, show that CGSVM+Chi-square yields the best sensitivity of 81.4%. In Scenario 2, LSVM+ANOVA yields the best sensitivity of 80.5% compared to the rest of the models, and in Scenario 3, FGSVM+BACET gives the best sensitivity of 72.3% compared to the rest of the models.
Conclusion: Overall, the study demonstrates that the radiomics and feature selection methods employed in combination with the different support vector classifiers performed significantly and achieved decent results for the classification of CT pulmonary nodules. The outcome thus can help the clinicians to diagnose, and make better decisions and treatments.