machine learning based system for classification and detection of lung cancer

Abstract

newline Abstract newlineGlobally, lung cancer has the highest death rate among cancers, being responsible for the newlinemajority of cancer-related deaths. Identification of the lung cancer tissues or nodules is aided newlineby early diagnosis using imaging techniques. X-rays, CT, PET, MRI, and histopathological newlinescans (biopsies) are necessary for the screening of lung cancer. Computed Tomography (CT) newlinescan slices are preferred for finding nodule size and locations. However, handling these slices newlineis tedious and time-consuming for radiologists due to the manual diagnosis of scans (approximately newline250 300 slices per patient) and the requirement to distinguish between malignant and newlinebenign cells by advanced tests. newlineThe development of CAD systems to aid radiologists in lung nodule detection has made the newlinescreening process more feasible and reliable. Moreover, complex and large clinical data can newlinebe processed by CAD systems in the medical field. This allows them to gradually improve newlinetheir diagnostic performance by creating novel insights from data. Traditional CAD systems newlineinitially classified lung cancer through machine learning methods. However, a significant newlineamount of manual pre-processing is needed for these methods. Deep learning has proven to be newlinean excellent way to improve diagnosis, lessen the workload for radiologists, and help doctors newlinemake specific treatment decisions. newlineThe present thesis addresses the CAD-based lung cancer classification system, which includes newline(a) Pre-processing of Computed Tomography (CT) scan lung images, (b) a feature newlineextraction and dimensionality reduction algorithm for a machine learning-based classification newlineapproach, (c) an efficient framework for binary classification using transfer learning that operates newlineat high speed and without annotations, and (d) a color transformation-based deep learning newlineframework for multi-class classification of histopathological NSCLC sub-types. During the newlinepre-processing of CT scan images, the lung area was segmented using morphological operations newlineand K-means clustering (based on cos