Detection and Classification of Malignancy in Lung CT Images using Deep Learning and Hybrid Neural Network Technique

Abstract

Lung cancer-is-the foremost motive of-death-and is most diagnosed worldwide, both in men newlineand women. The existence of the pulmonary nodule is a possible symptom of lung cancer, newlineespecially since few stages are diagnosed at the initial step. If these nodules are detected at the newlinelocalized phase, the probability of survival can improve. Presently computed tomography is newlineoperated on to diagnose and treat lung-cancer. However, this CT scanner makes-a lot of data; newlineinterpretation and manual segmentation are difficult and time-consuming. It encumbers the newlineradiologist and increases their workload, and the probability of overlooking some pathological newlineelements, such as irregularities, are additional in a visual examination. The present research newlineaims to contribute to developing computer-aided-diagnosis and detection-of-lung cancer to newlineassist radiologists in efficiently diagnosing and detecting lung cancer. In particular, the current newlinework has been concentrated on developing the algorithmic component of image enhancement, newlinepulmonary nodule segmentation, nodule classification and malignant nodule segmentation for newlinecancer stage identification. newlineIn the medical image-capturing process, noise will exist in images, and proper enhancement newlineand pre-processing are required to analyze these images. Most researchers considered the same newlineon CT lung images using ROI selection, morphological operations, histogram equalization, newlineand binary thresholding methods and achieved around 95% accuracy. In the pre-processing newlinestage: Resampling, morphological closure, and image-denoising techniques were applied for newlinebetter accuracy. In the image segmentation stage: the LIDC dataset is used for labelled nodule newlineregions and the KDSB17 dataset for cancer/non-cancer labels and used Multi-Scale-LoG filters newlinewith boundaries of-area dimensions and U-Net models. High-level elements were pulled from newlinethe KDSB-2017 dataset and rebuilt as a 3D-Array operating a GMCAE. The candidates were newlinealso classified in 3D-deep-CNN and achieved 74% accuracy-with-a-validation-loss-of 0.57. newlineOn