Chronic obstructive pulmonary disease (COPD), a common lung disease is caused due to damages in small air sacs at the end of lung airways that affects both genders. Detecting and treating pulmonary emphysema at earlier stage can reduce the mortality. The lung Computer Tomography (CT) images obtained from Computed Tomography Emphysema Database (CTED) is utilized for emphysema classification. In this paper, the automated Neural Networks (NN) based Pulmonary Emphysema Classification System (PECS) is proposed to identify and classify the emphysema lung diseases using three phases. The preprocessing phase employs the Contrast Limited Adaptive Histogram Equalization (CLAHE) technique for contrast enhancement of images. Then the image feature extraction phase uses the Laws Local Binary Pattern (L2BP) for extracting the texture descriptor images. Finally, the classification is done using two stage Neural Network classifiers (NN-1 and NN-2) to classify the emphysema images. The NN-1 classifier is used to classify the as Normal Tissue (NT), or abnormal (CLE & PLE) class and then NN-2 classifier identify the abnormal CT images into either CLE or PLE.
Research Article
AN EFFICIENT HYBRID TEXTURE DESCRIPTOR FOR PULMONARY EMPHYSEMA CLASSIFICATION
https://doi.org/10.21203/rs.3.rs-1749811/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Pulmonary Emphysema Classification
Texture Descriptors
Local Binary Pattern
Laws Textures
Neural Network Classifier
Lung CT image analysis for the diagnosis of COPD is the recent CAD systems explored. Emphysema, a progressive long-term lung disease, may be detected using CT images of the lungs. Shortness of breath is one of the symptoms of this condition. Chronic bronchitis, airflow restriction, and emphysema are also symptoms of COPD. Exposure to cigarette smoke (the most frequent risk factor in Western nations) and smoke from the burning of biomass fuels within dwellings for heating or cooking are well-known risk factors for the development of COPD (the most common risk factor worldwide). Breathlessness, a persistent cough, sputum production, and wheezing are the most common symptoms of COPD, but patients may also feel ankle inflammation, tiredness, and chest discomfort. Although pathological criteria are used to identify emphysema, it is becoming more widely accepted that CT imaging techniques are the best way to diagnose the disease in vivo.
A novel LBP texture descriptor with LAWS texture descriptor is created using CT images. Emphysemas can be seen on CT images as low-attenuation regions near the airways. Normal Tissue (NT), PanLobular Emphysema (PLE), ParaSeptal Emphysema (PSE), and CentriLobular Emphysema (CLE), are the different forms of the emphysema diseases. Lung CT images are obtained from CTED for emphysema classification is shown in Figure 1.
CT image analysis using the rapid texture measure method for the automated NN classification system is presented in [1]. Each pixel of a textured image aids the detection of lesion region in the CT images. The texture energy method generally takes minimum 5x5 convolutions. Rotation invariance method is used to texture energy features. Automated NN based glaucoma diagnosis model is developed in [2] using the micro statistical descriptors. Initially, ROI extraction with micro textures feature extraction from the micro-statistical descriptors is done. Finally, classification using linear vector quantizer based Artificial Neural Network (ANN).
Various texture feature measures for image classification are comparatively studied in [3]. Among all the features, gray-scale invariant approach utilizes the LBP feature descriptors to extract and evaluate the texture measurements. Gray-scale indifference is a property of LBP and requires large sample size. Center-symmetric covariance features outperformed Laws' measurements for several images. Emphysema classification analysis using LBP feature extraction in lung CT image is demonstrated in [4]. Instead of utilizing the whole image texture feature extraction, only six statistical data of uniform LBP texture descriptors with SVM classifier is performed. The optimization is done using the elephant herding optimizing algorithm.
Texture feature based classification for pattern recognition letters using the dominant rotated LBP feature extraction pattern is discussed in [5]. The method provides all of the structural information using LBP extraction, but it also collects additional data by employing magnitude data, resulting in greater discriminative power. The dominant rotated LBP descriptor approach aids to achieve rotation invariance. It is followed by feature selection for the lung CT image classification. Deep Learning (DL) based skin melanoma classification system using the Generalized Autoregressive Conditional Heteroscedasticity (GARCH) approach is presented in [6]. In the first phase, colour features are clustered using k-means clustering. The second phase utilizes the local statistical descriptors from GARCH and dominant rotated LBP approach. The third is feature reduction by t-test method. Finally, DL is used for classification.
Pulmonary emphysema classification using LBP quantitative analysis of lung CT images is discussed in [7]. Gaussian filter is used for preprocessing. The quantitative analysis of hybrid LBP descriptors and pixel intensity features are used for ROI extraction. A NN classifier is used for classification. Measurement of plasma biomarker inflammation among asymptomatic smokers and COPD patients are done in [8]. The individuals are classified by evaluating the proportion of emphysema lesion level physiologically and using CT images. This aids to differentiate COPD patients from others with lower chance of developing the disease.
Two-stage NN with convolutional layers for the lung nodule CT image detection is explained in [9]. First stage is the U-Net based detection network and the second stage is dual pooling structure mask by the data augmentation method. Multilayer NN using levenberg marquardt algorithm aids the diagnosis of COPD [10]. Back propagation with momentum approach is used to avoid similarity problem. Then the classification is done using multilayer NN with two hidden layers. Computerized artificial NN approach is described in [11] for diagnosis and classification of COPD diseases with four classes classification is performed.
Computer-aided emphysema classification diagnosis using the bio-inspired algorithm is presented in [12]. ROI feature extraction with spatial, fuzzy c-means clustering algorithm extracts the shape, texture, etc. Moth–flame optimization and firefly optimization are used for feature selection. Extreme machine learning with SVM classifier is employed. Lung disease classification using the robust texture feature extraction is discussed in [13]. Local quinary patterns, multifractal and intensity histograms features are all fused together. Before employing an SVM classifier, an auto-encoder network and principal components analysis are utilized to reduce the feature space vector for emphysema disease classification.
Higuchi’s dimension and multi-fractal techniques for emphysema image pattern detection is presented in [14]. For detecting emphysema pattern lesion region, the first technique uses several Higuchi dimension algorithm. The second method is to extract multi-fractal spectrum in emphysema pattern identification and analysis using High-Resolution CT (HRCT) images. Emphysema classification for ensemble features extraction from global average pooling layer of CNN network is explained in [15]. This utilizes the ImageNet, with SVM based error correcting output code classifier for emphysema disease classification.
Computerized analysis of COPD using DL approach is designed in [16]. Hilbert-Huang transform based statistical features are extracted from the CT images. Then DL algorithm with CNN [17] model classifier is used for classification of pulmonary emphysema. Lesion margin feature extraction approach for lung CT image classification is discussed in [18]. Morphological features such as convexity, circularity, etc. are extracted using the Otsu thresholding method. Then the classification is done with Multi-layer perception classifier [19].
In most cases, are intensity based detection methods using the textural features, invariant moment, and other factors. The classification task in this paper is to identify the subclasses of emphysema diseases. Figure 2 shows the PECS system for COPD lung emphysema detection and classification.
The PECS consists of three modules; preprocessing, features extraction, and classification as in many pattern recognition system.
3.1 Preprocessing Module
In the first module of PECS, the contrast of the CT image is enhanced by using Contrast Limited Adaptive Histogram Equalization (CLAHE). It uses local intensity histogram for the enhancement and thus the over enhancement by conventional histogram equalization is eliminated.
3.1.1 Contrast Limited Adaptive Histogram Equalization (CLAHE):
CLAHE is a version of Adaptive Histogram Equalization (AHE) that corrects for contrast over amplification. It works on patterns, which are tiny areas of an image rather than the full image. To remove the false borders, the nearby pixels are merged using bilinear interpolation. The contrast of images may be improved with this technique. CLAHE may also be used on colour images, When only the luminance channel of an image is equalized, the results are much better than when all of the channels of an RGB image are equalized. When applying the CLAHE technique, two parameters are to be considered,
- ClipLimit – This option determines the contrast limiting threshold. The default value for this field is 40.
- TileGridSize – This specifies how many tiles are in each row and column. This is set to 8x8 by default and aids to add tiles to an image while tiles are split.
In CLAHE, the transformation slope function causes the contrast amplification in the vicinity of a certain pixel value. The neighbourhood slope cumulative distribution function is proportional to the value of the histogram at that pixel value. Before computing the distribution function, CLAHE restricts the amplification by clipping the histogram at a predetermined value. The distribution function slope and its transformation function, is limited and shown in Figure 3.
Figure 4 shows the conventional histogram equalization and CLAHE output of contrast enhanced lung CT images obtained from the preprocessing of PECS system. It shows the image of NT, CLE and PLE sample image fed into the preprocessing phase. Figure 5 shows the HE and CLAHE texture descriptor pattern obtained after preprocessing of the sample PSE image given in Figure 4(c).
3.2 Feature Extraction Module
In the second module of PECS, a new texture descriptor, Laws Local Binary Pattern (L2BP) which is a combination of Laws texture map [1], [2] and Local Binary pattern (LBP) [3], [4] is designed. The process of extracting higher-level information from an image, such as colour, shape, and texture, is known as feature extraction.
3.2.1 LBP Features
LBP, a Textural descriptor is used in this study to sort COPD CT images into several classes. [5]. LBP was initially introduced in 1994, and it has since been proven to be a powerful texture classification feature extraction technique. The detection performance on CT image datasets is considerably enhanced when LBP is coupled with the Histogram of Oriented Gradients (HOG) descriptor. LBP was first developed to assess local image contrast before being used as a statistical and structural texture descriptor.
Applying LBP approach to the image provides this histogram as a feature vector. The LBP feature vector is generated as follows:
- Divide the image that needs to be examined into cells and further into pixels such as 16x16 pixels for each cell.
- Compare each pixel in a cell to each of its eight neighboring pixel (on its left-top, left-middle, left-bottom, right-top, etc.). Follow the pixels in a clockwise or counter-clockwise manner around a circle.
- Write "0" where the value of the centre pixel is larger than the neighbor's value. If not, write "1." This yields a binary number of eight digits.
- Calculate the histogram of the probability for each "number" occurring in each cell (i.e., each combination of the pixels that are smaller and that are greater than the center).
In this paper, the feature vectors may now be processed using a NN, an ensemble learning network to classify images. Texture analysis of lung CT images is performed with LBP feature extraction for diagnosing emphysema and its subclasses using two stage NN classifiers.
Figure 6 shows the neighbourhood pixel size used to create the patterns that aid to determine the vector features.
Textons are the histograms of texture elements that are used by the LBP to define textures. The local structure is used to specify the texture for each pixel. The intensity level variations between neighborhood pixels are used to extract binary information. Binary information is extracted using the intensity level differences between neighboring pixels. A pixel's intensity level serves with the threshold value for pixels nearby. Reading values in certain directions to represent distinct spots, curving edges, etc., is referred to as binary coding.
LBP has shown to be an excellent texture descriptor with a low computational cost. Because of its excellent quality, it has attracted a lot of attention from researchers, and several adjustments have been proposed throughout time. The original LBP operator creates binary codes or patterns using an 8-neighborhood method, which requires eight pixels around the centre pixel. For texture descriptors, the original LBP's rotation variation isn't always a good thing. Simultaneous intensity fluctuations, on the other hand, have no effect on it.
When eqn. (1) is applied to the entire image, LBP with distinct image pattern is generated as shown in Figure 7.
For each colour channel, all of these images are extracted without regard for the border. This phase resulted in LBP texture features.
3.2.2 LAWS texture map approach
Laws' approach is a feature extraction technique that uses local texture energy measures. In order to extract the micro structural components of the image such as Ripple, Wave, Spot, Edge and Level, five matrices are defined. Laws vectors are shown in Table 1. Based on these vectors, 25 kernels are derived and they are convolved with the CT image to extract textures.
Table 1
Laws vectors
|
Micro Structures |
Matrices |
|
Ripple (R) |
[ l,-4, 6,-4, 1] |
|
Wave (W) |
[-1, 2, 0, -2, 1] |
|
Spot (S) |
[-1, 0, 2, 0,-1] |
|
Edge (E) |
[-l,-2, 0, 2, 1] |
|
Level (L) |
[ 1, 4, 6, 4, 1] |
The Laws kernels for extracting textures are shown in Table 2. Laws created five labeled vectors that may be used to create two-dimensional convolution kernels.
Table 2
Laws kernels for extracting textures
|
|
Ripple (R) |
Wave (W) |
Spot (S) |
Edge (E) |
Level (L) |
|
Ripple (R) |
RR |
RW |
RS |
RE |
RL |
|
Wave (W) |
WR |
WW |
WS |
WE |
WL |
|
Spot (S) |
SR |
SW |
SS |
SE |
SL |
|
Edge (E) |
ER |
EW |
ES |
EE |
EL |
|
Level (L) |
LR |
LW |
LS |
LE |
LL |
When these masks are convolved with a textured image, they extract specific structural elements. Texture is an important aspect of human vision. Regardless of window size, the texture energy conversion takes approximately the same amount of time.
By using the Laws kernels in Table 2, 25 new images are generated from one CT image having different texture measures. Instead of extracting one LBP from the raw CT image, it is extracted from the Laws texture map and thus 25 LAWS LBP's (L2BP) are extracted from the raw CT image as shown in Figure 8 and Figure 9. When image texture areas are known to be large, this would allow extremely large (highly dependable) texture energy windows to be used; however the number of buffered filtered image rows would also be large as depicted from Figure 8.
Normalization is used to modify the source image's brightness and contrast so that textures that differ only in luminance or contrast cannot be identified. Normalization is accomplished by removing the local average brightness from each texture energy row and dividing by the local average contrast. The Laws technique extracts secondary features from the image's intrinsic microstructure features using filter masks, which may subsequently provide L2BP feature extraction pattern for PSE pattern descriptor of Figure 5. This L2BP pattern features extracted are utilized for the emphysema subclass classification.
3.3 Classification Module
The third stage of the PECS classifies the CT image into NT, CLE, PSE or PLE. This stage uses NN classifier. Due to the larger optimization problem and high computational cost, the direct multi-class classification is not employed instead two different NN classifiers; NN-1 and NN-2 are designed. The former classifier classifies the input CT image into either NT or abnormal (CLE & PLE) and the later one predicts the abnormal CT images into either CLE or PLE.
3.3.1 Neural Network Classifier:
A neural network (NN) is a network or circuit of neurons made up of artificial neurons or nodes that was recently termed as an Artificial Neural Network (ANN) [17]. A NN is an artificial machine learning or deep learning [6] algorithm with image processing that is used to solve AI issues. In ANN, the connections of biological neurons are represented as weights between nodes. A positive weight represents an active link, whereas a negative weight represents a suppressive link. All inputs are given a weight before being combined together. This action is described as a linear combination. NN have been used to solve a wide range of issues. An ANN or Simulated Neural Network (SNN) with artificial neurons is an integrated set of artificial nodes linked into the convolutional layer that process data using a mathematical or computational framework on a contributed to the increasing method to computing.
An ANN is an adaptive system that changes its structure in response to data flowing through a network, whether external or internal, in most cases. Nonlinear statistical data analysis or decision-making tools in a more efficient manner are nonlinear statistical data analysis and decision-making tools. They can be used to analyze complicated input-output connections and to examine the patterns in images. In this module, NN approach is employed for image classification [19]. Deep learning model consists of more hidden layers between input and output layers with nonlinear relationships. Whereas, traditional a simple NN architecture normally have one hidden layer is shown in Figure 10.
NN may be applied in a variety of domains such as; Time series prediction and modeling are included in function approximation, or regression analysis. Image processing, signal processing, filtering, clustering, speech, handwritten, pattern, sequence, text recognition, medical diagnosis etc., are the various applications of NN.
In the proposed PECS model, the two stage NN classifiers (NN-1 and NN-2) are used. First, the NN-1 classifier is used to classify the images as either normal tissue (NT) or abnormal tissue (CLE & PLE/PSE), and then the NN-2 classifier is used to classify the abnormal CT images as either CLE or PLE or PSE.
The proposed systems are evaluated using Computed Tomography Emphysema Database (CTED) [7] images. The HRCT images are acquired from 39 subjects and consist of 115 images. Also, it provides ground truth data by a chest radiologist and a pulmonologist and classified into three classes; NT, CLE, and PSE or PLE. Table 3shows the description of CTED database used for COPD analysis. The proposed PECS system with LLBP method was implemented in Matlab R2016a tool for classification analysis.
|
Category |
#images |
Resolution |
|---|---|---|
|
NT |
59 |
61x61 pixels |
|
CLE |
50 |
|
|
PSE/PLE |
59 |
To test the proposed system, a re-sampling procedure named Cross-Validation (CV) is employed where a portion of data is used for training and later uses the remaining data for testing. The Neural Network training (NN-traintool) tool window for training the sample images for the proposed system is shown in Fig. 11.
As the number of data available to validate the model is limited, the k-fold technique in CV is used. The accuracy, sensitivity and specificity measures are used to evaluate the PECS model performance using 10-fold cross-validation applied to the whole training data set. Figure 12 shows the k-fold technique in the CV with k = 5. In this study, 10-fold CV is employed.
Table 4 shows the performance metrics used for PECS systems. When the system correctly identifies a positive result (abnormal), it is referred to True Positive (Tp) refers to when the system properly detects a positive result (abnormal) whereas, False Positive (Fp) refers to when the system mistakenly identifies a positive result. For recognizing negative (normal) findings, two more terms namely, True Negative (TN) and False Negative (FN) are referred.
Table 4
Performance metrics
The K-fold cross-validation is evaluated to analyze the performance of the proposed PECS system. Here k = 10 fold is considered for the cross-validation and to generate the confusion matrices for 1st stage and 2nd stage of the PECS classification system. Table 5 shows the LBP and LLBP pattern descriptor based accuracy measure (correctly tested sample images) and the evaluation of K = 10 fold cross-validation for 1st and 2nd stage PECS classification system.
|
K = 10 Fold Cross-Validation |
Accuracy (Correctly Classified Testing Sample images) |
|||
|---|---|---|---|---|
|
1st Stage Classification (NN-1 classifier) |
2nd Stage Classification (NN-2 classifier) |
|||
|
LBP Pattern (%) |
LLBP Pattern (%) |
LBP Pattern (%) |
LLBP pattern (%) |
|
|
Fold 1 |
82.4 |
76.5 |
90.9 |
100 |
|
Fold 2 |
100 |
100 |
81.8 |
90.9 |
|
Fold 3 |
76.5 |
70.6 |
90.9 |
100 |
|
Fold 4 |
76.5 |
76.5 |
100 |
100 |
|
Fold 5 |
94.1 |
94.1 |
100 |
100 |
|
Fold 6 |
87.5 |
68.8 |
90.9 |
100 |
|
Fold 7 |
76.5 |
82.4 |
100 |
100 |
|
Fold 8 |
100 |
93.8 |
100 |
100 |
|
Fold 9 |
76.5 |
70.6 |
100 |
100 |
|
Fold 10 |
58.8 |
82.4 |
81.8 |
90.9 |
Table 5 depicts the 10-fold cross-validation accuracy performance measure for two stage classification. The confusion matrices and the ROC curve plot for both LBP and the proposed LLBP descriptors are simulated and discussed for PECS model.
4.1 NN -1 Classifier – First Stage Classification
In the 1st stage Neural Network classifier (NN-1), the testing samples of CT images are used to classify the pulmonary emphysema diseases into normal or abnormal images. The high resolution lung CT images are obtained from CTED database for emphysema classification. The classification results are evaluated using the 10-fold cross-validation and the corresponding confusion matrix and ROC curve plot is demonstrated in Fig. 13 and Fig. 14 respectively for both LBP and LLBP.
From Fig. 13, the 10 fold confusion matrix plot provides 95.8% of accuracy, 96.3% of sensitivity and 94.9% of specificity performance measures for proposed PECS model using LLBP descriptor.
4.2 NN -2 Classifier – Second Stage Classification
The abnormal emphysema sample image output obtained from first stage classifier is fed as input to the NN-2 classifier. In the 2nd stage classification, the abnormal image is further classified into CLE and PLE/PSE. The classification results are evaluated using the 10-fold cross-validation and the corresponding confusion matrix and ROC curve plot is demonstrated in Fig. 15 and Fig. 16 respectively for both LBP and LLBP.
From Fig. 15, the 10 fold confusion matrix plot with 98.2% of accuracy, 98.3% of sensitivity and 98.0% of specificity performance is measured for proposed PECS model using LLBP descriptor. From Fig. 15 and Fig. 13, the average performance measures for proposed LLBP pattern is provided with 97% of accuracy, 97.3% of sensitivity and 96.45% of specificity. Table 6 predicts the performance measure comparison of existing and proposed method.
|
S. No |
Author Details |
CT Lung Images Database |
Methodology Used |
Accuracy Measure |
|---|---|---|---|---|
|
1. |
Tuba, E., et al. [4] |
Computed Tomography Emphysema Database (CTED) |
Uniform LBP pattern for statistical feature extraction, SVM classifier with Elephant Herding Optimization |
Accuracy (89.72%) |
|
2. |
Sørensen, L., et al. [7] |
Department of Respiratory Medicine, Gentofte University Hospital |
Gaussian-based filter for preprocessing, ROI, Textural features, LBP for Quantitative analysis, Nearest Neighbor classifier. |
Accuracy (95.2%) |
|
3. |
Er, O., et al. [9] |
Chest Diseases Hospital’s Database (own database) |
Levenberg–Marquardt (LM) algorithm, Back propagation with momentum for features, Multilayer Neural Networks with two hidden layer for classification. |
Accuracy (95.43%) |
|
4. |
Isaac, A., et al. [12] |
Real-time and Public Dataset |
ROI feature extraction, Spatial Intuitionistic Fuzzy C-Means clustering algorithm, shape, texture, and run-length features, Moth–Flame Optimization, Firefly Optimization for feature selection, SVM based Extreme Machine Learning classifier. |
Accuracy (95.43%) |
|
5. |
Bhuma, C. M. [15] |
Computed Tomography Emphysema dataset (CTED) Images |
Ensemble Features extracted from global average pooling layer of CNN with ImageNet, Error Correcting Output Code classifier. |
Accuracy (95.88%) |
|
6. |
Altan, G., et al. [16] |
Public Dataset |
Hilbert-Huang transform based statistical features, Deep learning algorithm with CNN model classifier. |
Accuracy (93.67%) |
|
7. |
Riti, Y.F. et al. [18] |
DR.Sardjito Hospital Yogyakarta as image dataset |
Otsu Thresholding, Morphological features, convexity, solidity, circularity, and compactness, Multi Layer Perception (MLP) classifier. |
Accuracy (85%) |
|
8. |
Proposed System |
Computed Tomography Emphysema dataset (CTED) images |
Pulmonary Emphysema Classification System (PECS), Contrast Limited Adaptive Histogram Equalization (CLAHE) technique for image contrast enhancement, Laws Local Binary Pattern (L2BP) descriptor feature extraction (combination of LBP and LAWS pattern), Two stage Neural Network classifiers (NN-1 and NN-2) |
Average Accuracy of 1st and 2nd stage (97%) |
Table 6 predicts the performance measure comparison of existing methodologies with the proposed PECS model methodology for pulmonary emphysema disease classification. Thus the accuracy measure is predicted high with maximum correctly classified test images using PECS classification system.
In this study, the PECS proposed system is presented to diagnose the emphysema disease subclasses using lung CT images obtained from CETD database. Multiple features by LBP and LAWs texture features are extracted from CT image patterns. LAWS texture pattern and LBP texture descriptors are combined together to form the LLBP descriptor pattern. The two stage artificial NN-1 and NN-2 classifiers are used for pulmonary emphysema disease detection and classification into CLE and PLE or PSE subclasses using PECS classification model. Various characteristics collected from emphysema features of are utilized for training the NN model and then test samples are given as input for NN classifiers. The classification results are obtained and the model is evaluated using the 10-fold cross-validation. The corresponding confusion matrix and ROC curve plot is demonstrated for proposed LLBP feature descriptor with average accuracy of 97% is obtained for both stage one and stage two classifications.
CONFLICT OF INTEREST STATEMENT:
On behalf of all authors, the corresponding author states that there is no conflict of interest.
- Laws, K. I, 1980, "Rapid texture identification", Image proc. for missile guidance, vol. 238, pp. 376-381.
- Mukil Alagirisamy. Micro Statistical Descriptors for Glaucoma Diagnosis Using Neural Networks. International Journal of Advances in Signal And Image Sciences, 7(1), 1-10, 2021.
- T. Ojala, M. Pietikäinen and D. Harwood, “A comparative study of texture measures with classification based on feature distributions,” Pattern Recognition, vol. 29, no. 1, pp. 51-59, 1996.
- E. Tuba, I. Strumberger, N. Bacanin, M. Tuba, Analysis of local binary pattern for emphysema classification in lung CT image, IEEE International Conference on Electronics, Computers and Artificial Intelligence (ECAI), 2019.
- R. Mehta and K. Egiazarian, “Dominant rotated local binary patterns (DRLBP) for texture classification,” Pattern Recognition Letters, vol. 71, pp.16-22, 2016.
- R. Thamizhamuthu and D. Manjula, "Skin melanoma classification system using deep learning," Computers, Materials & Continua, vol. 68, no.1, pp. 1147–1160, 2021.
- L. Sørensen, S. B. Shaker, and M. de Bruijne, Quantitative Analysis of Pulmonary Emphysema using Local Binary Patterns, IEEE Transactions on Medical Imaging 29(2): 559-569, 2010.
- S. B. Shaker, K. A. von Wachenfeldt, S. Larsson, I. Mile, S. Persdotter, M. Dahlbäck, P. Broberg, B. Stoel, K. S. Bach, M. Hestad, T. E. Fehniger, and A. Dirksen, Identification of patients with chronic obstructive pulmonary disease (COPD) by measurement of plasma biomarkers, The Clinical Respiratory Journal 2(1): 17-25, 2008.
- Cao, H., Liu, H., Song, E., Ma, G., Xu, X., Jin, R., ... & Hung, C. C. (2020). Two-stage convolutional neural networks for lung nodule detection. IEEE journal of biomedical and health informatics, 24(7), 2006-2015.
- Er, O., & Temurtas, F. (2008). A study on chronic obstructive pulmonary disease diagnosis using multilayer neural networks. Journal of medical systems, 32(5), 429-432.
- Işık, Ü., Güven, A., & Büyükoğlan, H. (2015, October). Chronic obstructive pulmonary disease classification with artificial neural networks. In 2015 Medical Technologies National Conference (TIPTEKNO) (pp. 1-4). IEEE.
- A. Isaac, H. K. Nehemiah, A. Isaac, A. Kannan, Computer-Aided Diagnosis system for diagnosis of pulmonary emphysema using bio-inspired algorithms, Computers in Biology and Medicine, 2020.
- H. Li, R. Mukundan, Robust Texture Features For Emphysema Classification In CT Images, European Signal Processing Conference (EUSIPCO), 2020
- I. M. Adekunle, A. Funsho, New Approach for Emphysema Pattern Detection in Computed Tomography Images, Nigerian Online Journal of Educational Sciences and Technology 3(1): 9-17, 2021
- C. M. Bhuma, An Ensemble Approach To Emphysema Classification, IEEE International Conference on Industrial and Information Systems (ICIIS), 2020
- Altan, G., Kutlu, Y., & Allahverdi, N. (2019). Deep learning on computerized analysis of chronic obstructive pulmonary disease. IEEE journal of biomedical and health informatics, 24(5), 1344-1350.
- Murugan, S., Mohan Kumar, S., & Ganesh Babu, T. R. (2020). CNN model Channel Separation for glaucoma Color Spectral Detection. International Journal of MC Square Scientific Research, 12(2), 1-10.
- Riti, Y.F., Nugroho, H.A., Wibirama, S., Windarta, B. and Choridah, L., 2016, August. Feature extraction for lesion margin characteristic classification from CT Image lungs image. In 2016 1st International Conference on Information Technology, Information Systems and Electrical Engineering (ICITISEE) (pp. 54-58). IEEE.
- Shahada, S.A.A., Hreiji, S.M., Atudu, S.I. and Shamsudheen, S., 2019. Multilayer Neural Network Based Fall Alert System Using IOT. International Journal of MC Square Scientific Research, 11(4), pp.1-15.
No competing interests reported.