Hybrid Model for Lung Nodule Segmentation based on Support Vector Machine and k-Nearest Neighbor

The synergy of recently developed diagnostic radiology and machine learning algorithms has assured far reaching implications for the healthcare industry. At present, radiologists have access to top notch computer aided diagnostic (CAD) systems to create a consequence of the amplifying use and substantial applications of AI tools built right on the top of simple machine learning algorithms. This article proposes a model that extracts lung nodules from a 2 dimensional computed tomography (CT) slice by utilizing synthetic minority over-sampling technique (S MOTE) along with support vector machine (SVM) and k-nearest neighbor (K-NN) on a dataset of SPIE-AAPM Lung CT Challenge, 2015. Morphological transformations were performed on the 2D CT slices to achieve lung segmentation. Shape and textural features were retrieved into a vector to represent the region of interests (ROIs) from the lungs. Further, SMOTE was applied to resolve the issue of an imbalanced training data set which had very few samples of positive class in comparison with the samples of negative class. This ensured unbiased training of the classifiers and higher sensitivity towards the positive class. In the proposed work, two binary classifiers are combined in order to get an efficient model that exploited the individuality of both the classifiers. First, SVM and k-NN are trained separately on the balanced training dataset and then the outputs of both the classifiers are combined using simple sum rule to make the final prediction based on the collective scores for each data sample. Consequently, the resultant predictions depend on the collective performance of both classifiers for enhancing the overall efficiency of the model. The proposed hybrid model of SVM-k-NN outperforms the individual models with a sensitivity of 94.45% and G-Mean value of 94.19%. The model concentrates on accurately predicting the presence of a nodule and not for misclassifying a positive sample as it may lead to a huge loss to the patient.CCS CONCEPTS• Diagnostic radiology • computer aided diagnostic system (CAD) • machine learning