Development of Deep Learning Algorithms for Early Prediction of Lung Cancers

Abstract

As lung cancer is one of the leading causes of death, it is essential for early newlinedetection of cancer to increase survival probability. Computed Tomography (CT) newlinescan is the most preferred choice for the early screening and detecting lung newlinediseases. However, as an advanced CT scanner produces a large volume of CT newlinescans, manual diagnosis and marking of lung diseases are a time consuming and newlinelaborious task even for experienced Radiologists. A practical solution to this is an newlineautomatic Computer-Aided Diagnosis (CAD) of lung CT scan to assist the newlineRadiologists. This thesis mainly focuses on developing deep learning algorithm newlinebased approaches for the early prediction of lung cancers, which is carried out in newlinethree phases (a) Lung Segmentation, (b) Interstitial Lung Disease (ILD) newlineClassification and (c) Lung Cancer Classification. Automatic diagnosis of the lung newlineCT scan is a multi-stage process in which accurate lung segmentation is essential newlineto the robustness of the entire CAD system to diagnose lung diseases. However, newlinethe development of automatic lung segmentation becomes complicated in the newlinepresence of dense abnormalities formed by ILD or Lung cancer. In this thesis, a newlineconditional Generative Adversarial Network (c-GAN) based approach is developed newlineto eand#8629;ectively segment the lung region from the surrounding chest region. In the newlineproposed segmentation algorithm, the given lung CT slices are passed through the newlinetrail of encoders which encode these slices into a set of feature maps. Further, a newlinemulti-scale feature extraction module is designed, which extracts multi-scale newlinefeatures from the set of encoded feature maps. Finally, the decoders are used to newlineobtain the lung segmentation from the multi-scale features. The Multi-Scale newlineFeature Extraction (MSFE) makes the network learn the relevant features of dense newlineabnormalities. In contrast, the iterative down-sampling followed by the newlineup-sampling makes it invariant to the size of the dense abnormality. A Taguchi newlineiv newlinebased approach is demonstrated for selecting a suitable architecture of c-GAN