Investigation in lung ct scan for better classification of tumors using low complex features and deep learning schemes

Abstract

Lungs are the most valuable and important part of the human body since it produces newlinevital ingredients to different parts of the body. High incidence of lung cancer is due to newlineexcessive use of tobacco or alcohol followed by pollution and smoking. Lung cancer is newlinemost commonly seen in both men and women due to changes in living factors. The newlinemortality rate of women due to lung cancer and is almost equal to death caused due to newlinecervical cancer and still increasing. The global cancer burden continues to rise newlinedramatically in developing countries. Lung cancer is clinically evident only in later newlinestages as its symptoms are related to pneumonia and swelling. Non-small cell lung newlinecancerand small cell lung cancers are common lung cancers. Computed tomography newlineprovides better visibility to tumor which could be improve the diagnostic procedure. newlineBetter automated diagnosis can be provided with precise segmentation of nodule and newlinefeature extraction phase.The results show that the extracted features could able to capture the textural integrity newlineand heterogenous nature of the tumor. The inherent characteristics could help in newlinedifferentiating benign and malignant tissues further, they could help in differentiating the newlinecategories of malignancy. The effectiveness of the classification algorithm in newlinedifferentiating normal and abnormal tissues is evaluated using accuracy, sensitivity and newlinespecificity. CANFIS is found to have 95.8%, 94.8%, and 95.9% respectively. The newlineperformance of the CANFIS algorithm is found to be affected by the membership newlinefunction. Precise definition of fuzzy values could improve the performance of CANFIS. newlineDue to the increased dimension of feature vector Genetic Algorithm (GA) is employed newlinewhile evaluating RF algorithm. It is observed that RF perform better with GA compared newlineto its performance without GA. RF with GA is found to have 97.1% accuracy, 96.1% newlinesensitivity and 96.2% sensitivity values compared to 94.1%, 94.1% and 94.3% of RF newlinewithout GA.