Author Email: email@example.com
According to the study of World Health Organization (WHO), lung cancer is leading cause of death among all types of cancers (L. A. Torre et al., 2015). Lung cancer detection in early stage is became easy with the help of several image processing and machine learning techniques. Artificial Neural Network is proven to be the best technique for medical imaging, especially in lung cancer diagnosis. For lung nodule detection Computer Tomography (CT) scan images are preferred by so many researchers. The lung CT scans are extracted from Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) (Armato III et al., 2015) dataset. To detect lung nodules from CT scans, various image preprocessing techniques such as thresholding, clearing border and some morphological operations such as erosion, closing and opening are applied and after detecting Region of Interest (ROI), Convolutional Neural Network (CNN) is used to classify lung nodules using preprocessed images and it has given accuracy as 91.66% and sensitivity as 83%.
In 2015, World Health Organization (WHO) has a survey for deaths and it was proved that cancer is second most leading cause of death (L. A. Torre et al., 2015). Lung cancer is most common cancer type which starts in lungs and can spread to other body parts. The human body has habit of checking and maintaining the growth of cells. Uncontrolled division of cells caused due to unbalanced system of growth of cells and it also forms mass. And this mass is nothing but tumor. Tumor can be produced in any body part and if it forms in lungs then it is known as lung nodules (I. Sluimer et al., 2006). Initially, lung nodules are very small in size (I. Sluimer et al., 2006).
Detection and classification of lung nodules at early stage is necessary in lung cancer diagnosis. Medical imaging is very useful in diagnosis of any kind of disease. Lung cancer diagnosis uses several kinds of images but CT scans are more preferred by previously implemented Computer Aided Diagnosis (CAD) (I. Sluimer et al., 2006) systems because it is more accurate (Armato III et al., 2015) and easy to process (Armato III et al., 2015). Recent studies have introduced different deep learning techniques for lung cancer diagnosis. Some of the studies are compared in literature review section. Most of the researchers have used LIDC-IDRI dataset (Armato III et al., 2015) to collect lung CT scan images. This dataset contains 1018 helical thoracic CT scans. Theses CT scans are taken from 1010 patients. These CT scan images are annotated by four trained radiologists. These annotation are according to the size of lung nodules and also the malignancy of cancer is given by these radiologists. This annotations are used to train the model. This study is divided into two phases: i) lung nodule detection also known as, Region of Interest (ROI) detection and ii) Lung nodule classification. ROI is the area where lung nodule like structure is present. For detection of ROI, some image processing techniques are applied on lung CT scan images as discussed further. These techniques includes thresholding, border clearing morphological erosion, morphological closing and morphological opening. After ROI detection, Convolutional Neural Network (CNN) is applied on preprocessed images to classify lung nodules and non-nodules. CNNs plays very important and efficient role in medical imaging. These networks are very good at feature extraction and classification of the images. The results of preprocessing and classification are discussed in results and analysis section. Lung nodule classification using CNN after preprocessing gives 91.66% accuracy and 83% sensitivity.
The purpose of conducting this study is to analyze the effectiveness of convolutional neural network (CNN) in lung nodule detection and classification. By going through various research papers it is discovered that neural networks has played an very important role for feature extraction and classification rather than other traditional approaches of image classification. Preprocessing of images intend to better accuracy instead of providing whole image (Hongyang Jiang et al., 2017), as it reduce false positive rate. We have performed feature extraction and classification of lung nodules with 91.66% accuracy. The future work will be to classify the lung nodules into benign and malignant forms (Tizita Nesibu Shewaye et al., 2016) to detect the malignancy level of cancer in early stage and survival rate can be increased.
386 total views, 1 views today