Automated detection of lung cancer using statistical and morphological image processing techniques

Abstract

Lung cancer represents the second most commonly diagnosed cancer among Jordanian population. Evidence that early detection of lung cancer may allow for more timely therapeutic intervention has provided the momentum for lung cancer screening programs around the world. In this study, a computer aided detection (CAD) system is proposed in an attempt to detect the lung cancer areas using computed tomography (CT) images. It is implemented as a "second reader" to help radiologists focus their attention on regions that might be missed during visual interpretation. The proposed CAD system has three main stages; Segmentation by thresholding the CT images, labeling the founded regions and then extracting some diagnostic features of each region for further analysis and interpretation. The study is trained, tested, and validated using images obtained from forty five patients. The obtained results perfectly match the radiologist's diagnosis in detecting the defected areas and quantitatively measuring its size, location, borders as well as displaying its other diagnostic characteristics. Moreover, the proposed system can detect misclassified regions.

Untitled

“…[6][7][8][9][10][11][12][13][14] to access a large number of CT Scan images. In addition to the public databases, Early Lung Cancer Action Program (ELCAP), the National Biomedical Imaging Archive (NBIA) [15][16][17], National Lung Screening Trial (NLST) [18], King Hussein Cancer Center [19], Apollo Specialty Hospitals, Chennai [20], Cornell University database [21], University of Michigan (Dept. of Radiology) database [22,23], Japanese University and American University data collection databases [24], Nelson Trial Database [25], Nagano lung cancer screening [26], University of Toronto [23], Brigham and Women's Hospital datasets [23], ACSC database [27], and University Medical Center Utrecht [28] employed CT Scan images for tasks of CADx.…”

“…Segmentation has significant importance by increasing accuracy, precision, reliability, and reducing computational cost for lung cancer detection [44]. Thresholding technique is extensively used in literature for segmentation by minimizing the variance within class and maximizing the variance between classes [9,[18][19][20]24,26,27,30,32,35,37,40]. To segment lung internal structure to separate the possible nodule and simplify the examination of lung region, multiple segmentation techniques comprise morphological operations [9], 3D morphological processing [36], morphological closing, and labeling [15,45], marker based watershed [6,7,33,43], histogram thresholding [30], detection method and thresholding based on Otsu's method [11], Gray level thresholding [25], K-means clustering with threshold ratio of R = 1 [46], fuzzy clustering method [41], deformable model based segmentation, shape model and edge model [47], fuzzy c-means, fuzzy-possibilistic c-means and weighted fuzzy-possibilistic c-means clustering algorithms [8], selective nodule and vessel enhancement filters [10], anatomical segmentation technique [14], region growing technique [38,43], sobel edge linear combination of Gaussians (LCG) [38], connected component analysis, vascular subtraction, and pleural surface removal segmentation techniques [36] are proposed in literature.…”

“…Some researchers combine this component with segmentation. Several algorithms, for example, connected component labeling [19], morphological operation, canny edge detection [18], PCA [30], marker controlled watershed [27], seed point detection, cluster formation, cluster merging, candidate location adjustment [10], Markov-Gibbs random field based appearance model, appearance-based registration [38], nodule volume determination, doubling time estimation, and VGI computation techniques [36] are used to remove extra muscle, border enhancement, and boundary extraction of lung walls, extract missing or overlapped regions.…”

“…After feature extraction, some relevant features are selected with high discrimination ability for classification, and other features are discarded [48]. To detect primary lung cancer, some features are extracted, for example area of the region, radius, length, width, the morphological feature of its boundary, and spatial histogram [19]. Similarly, area, perimeter, and eccentricity [7,33], structural features (area, convex area, equivalent diameter, and solidity) [18,34] and texture features (energy, mean, standard deviation) [18], and morphological and colorimeter features [6] are extracted features from the region of interest of lungs.…”

“…Similarly, area, perimeter, and eccentricity [7,33], structural features (area, convex area, equivalent diameter, and solidity) [18,34] and texture features (energy, mean, standard deviation) [18], and morphological and colorimeter features [6] are extracted features from the region of interest of lungs. Binarization is used to extract multiple features [16,17,20,27,30,31,39,47]. Several texture features proved to be useful a feature, including uniformity, entropy, maximum probability, inertia, inverse difference, correlation, homogeneity, dissimilarity, autocorrelation, cluster shade, cluster prominence, inverse difference normalized, sum entropy, sum average, sum of squares, sum variance, difference variance, difference entropy, information measures of correlation and maximal correlation coefficient extracted from gray level co-occurrence matrix (GLCM) [13,16,20,[30][31][32]39,47], sequential forward selection [28,31], spatial gray level dependence matrix (SGLDM) [40], genetic algorithm [32], masking approach [17,27], histogram [16,40], PCA [22,49], region growing technique [24,45], linear discriminant analysis (LDA) [15,45], filter bank method [28], box-counting method [40], contrast enhancement and calcification [34] and gray-we...…”