Order from us for quality, customized work in due time of your choice.
Abstract:
Lung cancer is one of the most serious cancers in the world, with the smallest survival rate after the diagnosis. In CT scans, lung nodules appear as dense masses of various shapes and sizes. They may be isolated from or attached to other structures such as blood vessels or the pleura. In this paper a detection of Candidate Nodules (solitary or juxtapleural in a 2D CT slice is achieved using two schemes of segmentation and enhancement algorithms. Convolutional Neural Network (CNN) for deep learning classification is used as a revolutionary image recognition method to distinguish between two types of nodules according to its location (juxtapleural and solitary lung nodules). Our CAD system achieves accuracies of first scheme of segmentation for detecting solitary nodules and juxtapleural nodules by using CNN are &&and &&. respectively. Also achieves accuracies of second scheme of segmentation for detecting solitary nodules and juxtapleural nodules by using CNN are &&and &&. respectively.
Keywords: CT scans, Convolutional Neural Network (CNN), Deep Learning, Computer Aided Detection (CAD).
I. Introduction
Lung cancer is a disease that consists of uncontrollable growth of cell and tissues of the lung which may lead to metastasis that is the infestation of adjacent and nearby tissue and infiltration beyond the lungs. From epithelial cells Carcinomas are derived which are the vast majority of primary lung cancers. Lung cancer, the most usual cause of cancer-imputed death in men and women. An estimated new lung cancer cases 14% for males and 12% females in US in 2017 [1].
The early detection of lung cancer can increase overall 5-year survival rates by extracting the lung nodules. Hence, this diagnosis can improve the effectiveness of treatment. Traditional x-ray and computed tomography (CT scan) are attempted to diagnose lung nodules. Treatment of lung nodules depends on the histological
type of cancer, the stage, and the patient’s status, but overall only 14% of people diagnosed with lung cancer survive five years after the diagnosis[2].
Because of small size of nodule in the lung, it is difficult to distinguish between it and another mass in a 2D slice.
Actually, the search for micro-nodules does not always make sense on single slices: the nodule shape, size and gray tone are very similar to vessels sections. therefore, a segmentation step is very important to distinguish between the small nodules and blood vessels. Hence, the type of nodules according to its location (solitary or juxtapleural) will be easily classified.
In this paper, we propose a two schemes of segmentation and enhancement for nodule emphasis which extracts a 2D candidate lung nodules. A convolutional neural network is used as a deep learning tool for the classification of juxtapleural and solitary lung nodules.
II. Related Work:
To date, many types of research about nodule detection by using CAD system have been developed. It begins with preprocessing and segmentation followed by the classification step.
For instance, Diego et al [3] used a method composed of four processes for lung nodule detection. The first step employed image acquisition and pre-processing. The second stage involved a 2D algorithm to affect every layer of a scan eliminating non-informative structures inside the lungs, and a 3D blob algorithm associated with a connectivity algorithm to select possible nodule shape candidates. The final step utilized a support vector machine for classifying the possible candidates into nodules and non-nodules depending on their features. QingZeng et al [4] proposed to employ, respectively, the convolution neural network (CNN), deep neural network (DNN), and stacked autoencoder (SAE). The experimental results show that the CNN network archived the best performance with an accuracy of 84.15%, sensitivity of 83.96%, and specificity of 84.32%, which has the best result among the three networks. Yang et al [5] described a 3D pulmonary nodule detection scheme utilizing MSCT images. This method segmented the initial nodule candidates first and extracted voxels features based on analysis eigenvalues of Hessian matrix. Then support vector machine (SVM) and decision rule are applied to categorize them into two sorts to remove FPs. Sarah et al [6] used Gaussian smoothing kernel for filtration which helps to reduce noise effects. Next, features such as sphericity, mean and variance of the gray level, elongation and border variation of potential nodules are extracted to classify detected nodules to malignant and benign groups. Fuzzy KNN is employed to classify potential nodules as non-nodule or nodule with different degree of malignancy.
Serhat et al. [7] used Genetic Cellular Neural Networks (G-CNN) for segmentation. Then, for each lung region, ROIs were specified with using the 8 directional search; +1 or -1 values were assigned to each voxel. A 3D template was created to find the nodule-like structures on the 3D ROI image. Convolution of the 3D ROI image with the proposed template strengthens the shapes that are similar to those of the template and it weakens the other ones.
Finally, fuzzy rule based thresholding was applied and the ROIs were found. The computer aided diagnosis (CAD) system achieved 100% sensitivity with 13.375 FPs per case when the nodule thickness was greater than or equal to 5.625 mm.
Jin et al [8] proposed convolution neural network as a classifier to detect the lung nodules. The system achieved 84.6% of accuracy, 86.7% of specificity and 82.5% of sensitivity.
Thomas et al [9] developed CAD system to detect and localize 60.1% of all the nodules with an average number of 2.1 (1.5%) false positives per slice. In addition, three different types of neural network structures for this CAD system are tested and compared. With 95% confidence we can conclude that deeper neural networks decrease the false positives significantly
Wu et al [10] combined a several common image processing techniques with complex segmentation step, such as connectivity labeling, binarization, mathematical operation and hole-filling.
Ming et al [11] applied the watershed algorithm to estimate the segmentation and then used a region growing method. Xujiong et al [12] proposed a two-step segmentation method for lung extraction. Firstly, a 3-D adaptive fuzzy thresholding technique and secondly applied a 2-D-based post refinement process on the lung contour chain code to obtain a complete lung mask. Michela et al [13] used a dynamic threshold for identification of three different groups corresponding, respectively, to the upper, middle and lower parts of the lung volume. The slices of the lung middle part and a threshold determined empirically for all other slices. In their tests, they applied thresholding with fixed threshold to the first 30 and the last 30 slices of a CT scan. Sasidhar et al [14] applied two steps of segmentation: firstly, extracted lung parenchyma by using a threshold of -420 Hounsfield Unit (HU) as:
Gray Level Value = 1024 + HU
Secondly, extract lung nodules using the threshold of -150 HU.
Qingxiang et al [15] proposed a method implements an active evolution and structure enhancement which can segment blood vessels and detect pulmonary nodules at a high accuracy.
Firstly, he introduced a vessel energy function (VEF) during active evolution to help distinguishing the nodules from vessels. VEF consists of three energy terms, which are gradient term, intensity term, and structure term.
= Fgradient û Fintensityû Fstructure
Secondly, he utilized a radius-variable sphere model to refine the extracted contours. Candidate blood vessel centerline points, denoted as
V = {,,&..,}, are first selected.
Serhat et al [16] segmented the lung regions of the CTs by using Genetic Cellular Neural Networks
(G-CNN). Then, for each lung region, ROIs were specified with using the 8 directional search; +1 or -1 values were assigned to each voxel. The 3D ROI image was obtained by combining all the 2-Dimensional (2D) ROI images. A 3D template was created to find the nodule-like structures on the 3D ROI image.
Qing et al [17] transformed a number of correlated variables into a smaller number of uncorrelated variables, which are called principal components depended on their cumulative variance proportion that is called principal component analysis (PCA).
III. Materials And Methods:
A. Dataset:
A 14 digital CT consisting of 2991 2D slices which contains 172 nodules (100 solitary nodule and 72 juxtapleural nodule) are collected with approval from Cornell University [18] each abnormal image contains a tumor with equivalent diameters of lung nodules ranging from 7.78 mm to 22.48 mm. The in-slice (x, y) resolution is 0.703×0.703 mm and the CT slice thickness is 1.25 mm in DICOM format and has 512×512 pixels.
B. Implementation:
- Software:
- Windows 10 Pro 64-bit
- Matlab R2018b toolboxes
- Hardware: