An Empirical Identification of Pulmonary Nodules using Deep Learning

Abstract

Cancer is among the top killers every year. Cancer is usually fatal, especially lung cancer. For detecting lung nodules quickly and correctly, preprocessing images by reducing noise, segmenting, and extracting interests is required since the lung CT scans used in this experiment are actual medical images. A median, Gaussian, Gabor, and Otsu algorithm were used in this research to evaluate lung regions. To diagnose lung cancer early, pulmonary nodules have been assessed for all visual characteristics. This helps Densenet201-GoogleLeNet achieve robust decisions regarding non-cancerous lung pixels and the implementation of linear, quadratic, cubic, fine-medium-coarse Gaussian SVM algorithms for lung pixels. The results with Gaussian were the best. The experiments findings indicated that the overall accuracy rate can be achieved under ten-fold cross-validation for all ranges is 100%, except for Densenet201, which is 99% and 97% in 70-30 and 90-10 ratios, and Googlenet, which is 98% for both 50-50 and 80-20 ratios. Detection rates of false positives should be zero. There has been an improvement in Densenet201's Matthews correlation coefficients compared to Googlenet's. In analyzing pulmonary CT images, the DG Network’s fusion SVM classifier algorithms provide 100% accuracy in detecting lung cancer. It is a feasible method to identify pulmonary nodules and can provide physicians with auxiliary diagnostic information. Thus, the author proposes that the above technique is the most appropriate for diagnosing lung nodules in an earlier stage.