Ensemble classification of integrated CT scan datasets in detecting … – Nature.com

Proposed methodology

The architecture of the proposed model as shown in Fig.1 considered CT scan images as the input to detect COVID-19 or non-COVID-19 images. The CT scan image datasets were collected and merged from three publicly available datasets. Since the dataset images were not of the same size, they were resized and merged. The images were then converted to grayscale from RGB. A modified region-based clustering method was proposed to segment the CT scan grayscale images. Furthermore, the model deliberated two feature extraction techniques including contourlet transform and CNN. Firstly, the contourlet transform method and secondly, the CNN feature extraction technique extracted feature vectors. These two vectors were fused in one feature vector, which was used as the input to train the classification model. The fused feature vector considered a large number of features that helped to accurately identify the COVID-19 or normal images. The system also proposed an authentic feature selection technique that extracted meta-heuristic features by using BDE. This optimized vector was subsequently used to recognize COVID-19 CT scan test pictures using an ensemble classifier.

Proposed methodology for detecting COVID-19.

The most important step in designing a computer-aided diagnostic (CAD) system for detecting COVID-19 at an early stage is the CT scan image segmentation22. In order to diagnose unusual disorders, segmentation is widely used in the area of medical images. Manual segmentation of the same medical images is possible. Image segmentation utilizing segmentation algorithms has a higher accuracy compared to manual segmentation. The original fuzzy c-means (FCM) algorithm23 works well for segmenting noise-free images, however, it fails to accurately segment the images with noise, outliers, or other imaging artifacts. The modified region-based clustering technique was used in this work to segment the CT images. The objective of the modified region-based clustering algorithm was updated to reduce the intensity of homogeneities by including spatial neighborhood information and altering the membership weighting of each cluster. The proposed segmentation algorithm has the following advantages: (a) propagates more homogeneous regions than other old fuzzy c-means algorithms, (b) manages noisy spots and (c) it is comparatively less sensitive to noise. These techniques have produced excellent output images with the simplest approach to isolate the objects from the background.

A chest CT scan is a useful medical imaging tool for accurately diagnosing COVID-19 cases24. As the open repository had a limited quantity of CT scan images, thus the images from all three databases were integrated to form a new database for this work. A total of 11,407 CT images with 7397 images from COVID-19 class and 4010 images from non-COVID19 class. The training and testing phases included images of COVID-19 and non-COVID-19.

The SARS-CoV-2 CT-scan dataset19 has 2482 CT scan images from 120 patients, including 1252 CT scans of 60 patients infected with SARS-CoV-2 from men (32) and females (28), and 1230 CT scan images of 60 patients who were not infected with SARS-CoV-2 but had other pulmonary disorders. The data of CT scan images was gathered from hospitals in Sao Paulo, Brazil. The CT scan images in this dataset are digital scans of printed CT tests, and there is no criterion for image size. The smallest CT scan images in the dataset are 324412 pixels, while the largest CT scans are 484456 pixels. In this dataset, the number of training and testing images are 1842 and 640 respectively.

The original CT scans image of 377 people are included in this COVID-19 CT image dataset20. There are 1558 and 4826 CT scan images, respectively, belonging to 95 affected COVID-19 people and 282 normal people. The Negin Medical Center in Sari, Iran, provided this dataset. All the CT image sizes are 2562563. In this dataset, the number of training and testing images are 5594 and 790 respectively.

These publicly available datasets are collected from authentic website21. This dataset contains a total of 2541 CT scan images with 1200 COVID-19 and 1341 non-COVID-19. In this dataset, a total of 1726 and 815 images are considered for the training and validation.

As the open repository had a limited quantity of CT scan images, the images from all three databases were integrated to form a new database for this work. A total of 11,407 CT images with 7397 images from the COVID-19 class and 4010 images from the non-COVID-19 class. Figure2 demonstrates sample CT scan images from each dataset. The training and testing phases included images of COVID-19 and non-COVID-19.

Sample CT scan images from three datasets.

Image pre-processing is a key step in medical image processing to obtain meaningful information and appropriate classification by eliminating noisy or distorted pixels from each CT scan image. In this stage, the images were first resized to 256256 pixels and transformed from RGB to grayscale images using the MATLAB function as the input for the model development. Color has no significance in detecting COVID-19 from the CT scan images hence grayscale images were employed during building the models to avoid any false classification and complexity. Grayscale images are simpler and easier to process than color images because they contain only one-color channel, which represents the intensity of the color for each pixel. Figure3 displays the preprocessing steps employed in this work.

Preprocessing steps applied to the COVID-19 and non-COVID-19 images.

Histogram equalization, an image processing technique that is frequently used on CT scan images to improve image quality in black and white color scales. The input images and its contrast-enhanced (after histogram equalization) images are shown in Fig.3 with the related histograms. Histogram equalization was achieved by efficiently spreading out the most frequent intensity values, extending the image intensity range. The adoption of a spatially variable histogram equalization technique seems to improve the visibility of anatomic structures in various clinical scenarios25. However, the technique increased the amount of noise and artifacts in the presented image.

The region-based clustering was employed to simplify the COVID-19 image region, which ensured less computational complexity and relatively accurate analysis. K-means, C-means, thresholding, morphology-based, edge-based, watershed, region-growing, and cluster-based approaches are among the various segmentation algorithms26. The authors of this paper proposed a cluster-based algorithm that segmented the image effectively and provided a better performance in terms of measuring evaluation matrices SSIM (structural similarity index), PSNR (peak signal to noise ratio) and RMSE (root mean square error) scores.

The proposed segmentation method partitioned the COVID-19 image into four clusters (C1 to C4) as gray matter (GM), cerebra-spinal fluid (CSF), white matter (WM), the necrotic focus of glioblastoma multiforme (GBM). The proposed segmentation technique employs an iterative process to locate the cluster region. In each iteration, the clusters centroid is modified to reduce the distance between pixels and the centroid. The mean brightness of all pixels within a cluster and the distance are obtained by using Eqs.(1) and (2) respectively. The COVID-19 segmentation process is depicted in Algorithm 1.

$${mu }_{k}= {C}_{k} sum_{i=0}^{N}frac{{Z}_{i}}{N},$$

(1)

$$r= left|{mu }_{k}-{x}_{i}right|,$$

(2)

where ({mu }_{k}) is the clusters mean intensity, and r means pixels distance from a clusters centroid. The intensity of the ith pixel within a cluster is ({Z}_{i}), ({C}_{k}) is the center of the kth cluster, and ({x}_{i}) is the intensity of the ith pixel. The number of pixels in a cluster is denoted by N. The COVID-19 segmentation process is depicted in Algorithm 1. Figure4 illustrates the grouping of COVID-19 image data step by step.

Applied modified region based clustering method for COVID19 and non-COVID19 image segmentation.

Algorithm 1: Proposed segmentation algorithm.

The contourlet transform tries to capture curves rather than points and includes anisotropy and directionality. The CT was created to solve the wavelet transforms limitations such as poor directionality, shift sensitivity and lack of phase information27. At each scale, it allows for a variable and elastic number of directions while obtaining virtually critical sampling. The contourlet transform28 is accomplished based on two steps including Laplacian pyramid decomposition and directional filter banks (DFB). At every level of the Laplacian pyramid, a down-sampled lowpass version of the source image is generated, as well as the difference between the source image and the down-sample lowpass image, resulting in a high-pass image. The next level Laplacian pyramid builds an iterative structure linking with the down-sampled lowpass version of the original signal. DFBs are used to create high-frequency sub-bands with a variety of directions. The contourlet transform acts on two-dimensional CT scan images. This work generated sixteen different multi-directional multiscale images using four-level CT with the 9-7 filter and computed thirteen various image features, including entropy, homogeneity, energy, correlation, and others from the segmented images, by enumerating the gray level co-occurrence matrix (GLCM) of each image. Figure5 presents the contourlet transformed images considering edges, lines, textures and contours in contrast to the wavelet transform.

Overall structure of contourlet transform feature extraction method.

For feature extraction, the proposed system employed the benchmark VGG19 CNN model, which outperformed the other CNN models such as AlexNet, GoogleNet, and ResNet50. A 19-layer version of VGGNet29 was used to create this network. Figure6 shows the VGG19 architecture, which includes sixteen convolution layers and three fully connected (dense) layers. For each convolution layers output, a non-linear ReLU was employed as an activation function. The entire convolution sections were divided into five sub-regions by five consecutive max-pooling layers. Two convolution layers were employed with depth dimensions of 64 and 128 respectively. Each of the other three sub-regions was made up of four consecutive convolution layers with depth sizes of 256, 512, and 512 in each sub-region. In this case, a convolutional kernel of size of 33 was chosen. The last layer of the proposed VGG19 models was replaced by a softmax classification layer. Two fully connected layers with neurons 1024 and 4096 were installed before the output layer. As a result, the fully connected layer yields 4096 features for classification.

Architecture of VGG19 for feature extraction from CT scan images.

A fusion-feature vector was created by combining the extracted features from the contourlet transform and CNN. Overlapping, redundancy, and dimensional expansion are regular occurrences in all fusion-based techniques, therefore dimension reduction, as well as redundancy minimization or the elimination of irrelevant features, is required to obtain the optimum features. Many researchers obtain optimized features using Principal Component Analysis (PCA)30 and minimum RedundancyMaximum Relevance (mRMR)31 but the BDE feature optimization method provides better performance than the others. For the dataset used in this study, three feature optimization approaches were tested and BED performed best.

In the mRMR feature selection algorithm, the mutual dependencies of x and y variable can be determined using Eq.(3) where p(x), p(y) and p(x,y) are the probability density functions.

$$Ileft(x,yright)= iint pleft(x,yright){text{log}}frac{p(x,y)}{p(x)p(y)}dxdy.$$

(3)

Equation(4) approximates the maximal relevance D(S,c), where xi is the mean of all mutual dependencies and c is the class. As a result, the function R(S), is represented by Eq.(5) that can be used to add minimal redundancies. S is the feature combination.

$${text{max}}Dleft(S,cright)= frac{1}{left|Sright|}sum_{{x}_{iin S, }}Ileft({x}_{i, }cright),$$

(4)

$${text{max}}Rleft(Sright)= frac{1}{{|S|}^{2}}sum_{{x}_{i}{{x}_{j}}_{in S, }}I({x}_{i, }{x}_{j, }).$$

(5)

In the PCA algorithm, the covariance of features is determined to take uncorrelated features. PCA uses Eq.(6) to combine the correlated features.

$$rho = frac{sum_{i=1}^{N}left({X}_{i}-overline{X }right)({Y}_{i}-overline{Y })}{n-1}.$$

(6)

The BDE feature selection technique is a heuristic evolutionary strategy for reducing the successive problem. The notion of advanced binary differential evolution (ABDE) is expanded to include feature selection difficulties. Three random vectors ({P}_{u1}), ({P}_{u2}), and ({P}_{u3}) are chosen for vector pk for the mutation operation, so that u1 (ne) u2 (ne) u3 (ne) k, where k is a population vector arrangement. The dth characteristic of the difference vector (Eq.(7)) is zero if the dth dimensions of the vectors ({P}_{u1}) and ({P}_{u2}) are equal; otherwise, it has the same value as the vector ({P}_{u1}):

$${difference, vector}_{k}^{d}= left{begin{array}{l}0, {P}_{u1}^{d}= {P}_{u2}^{d} \ {P}_{u1, other}end{array}right}.$$

(7)

Following that, the mutation and crossover processes are carried out, as illustrated by the Eqs.(8) and (9).

$${mute, vector}_{k}^{d}= left{begin{array}{l}1, {if, different ,vector}_{k}^{d}= 1 \ {{P}_{u3}^{d}}_{, other}end{array}right},$$

(8)

$${W}_{k}^{d}= left{begin{array}{l} {mute, vector}_{k}^{d} , if y le CR left|dright| d={d}_{random} \ {{P}_{k}^{d}}_{, other}end{array}right}.$$

(9)

Here, W denotes the try vector, ({CR}_{epsilon })(0, 1), a crossover amount, and ({gamma }_{varepsilon })(0, 1) denotes the mutation amount. If the try vector ({W}_{k}) has a higher fitness value than the current vector ({P}_{k}), then it will be replaced in the selection phase. In a different way, the current vector ({P}_{k}) is saved for the next generation. Finally, this fused method achieved 1300 accurate optimized features.

Figure7 illustrates the steps in obtaining the optimized features in a single vector by fusing the features vectors extracted by the contourlet transform and CNN. The size of this feature vector is 4109. BDE based feature selection method was then employed to get 1300 most discriminating features.

Block diagram of optimised feature selection process.

The authors suggested a novel, straightforward hybrid selective mean filter (HSMF) technique32 to calculate the average value selectively, unlike the traditional mean filter (MF) method, which calculates the average pixel utilizing all pixels in a given kernel region. A threshold value was used to define pixel selection (h). Noise was not considered in the noise reduction procedure if an adjacent pixel in a kernel was higher or smaller than the threshold value from the value of the core pixel. The pixel selection was performed with the following Eq.(10).

$${I}^{prime}left(x+i,y+jright)= left{begin{array}{l}Ileft(x+i,y+jright), quad if left|Ileft(x,yright)-I(x+i,y+j)right|le h\ 0, quad if left|Ileft(x,yright)-I(x+i,y+j)right|>hend{array}right..$$

(10)

If (left|Ileft(x,yright)-I(x+i,y+j)right|le h, for every i and j) then ({N}^{{{prime}}}left(x,yright)=N-1.) The noise image reduction is then calculated using Eq.(11).

$${I}_{SMF} left(x,yright)= frac{{sum }_{i=-frac{n-1}{2},j=-frac{m-1}{2}}^{+frac{n-1}{2},+frac{m-1}{2}}I^{prime}(x+i,y+j)}{N^{prime}(x,y)}.$$

(11)

In the Eqs.(10) and (11), the disparities between all nearby pixel values and the central pixel value are likely to exceed h in the edge areas. The pixel value ({I}_{SMF})(x, y) is equal to I in this situation (x, y). In contrast, in the homogenous regions, the disparities between all nearby pixel values and the central pixel value are likely to be smaller than h. The pixel value ({I}_{SMF}) (x, y) is equivalent to ({I}_{MF}) in such situations (x, y). Figure8 depicts the noise reduction process of the HSMF method. The mean pixel value at the central pixel in a position (x, y) was calculated only from the black area where the differences in pixel values from the value of the central pixel were less than the threshold value, not from all the pixels in a particular square kernel (i.e., union of black and red areas). The pixels outside of the black region, as well as those still inside the kernel of interest with pixel values higher than the threshold value, were not included in the calculation.

An illustration of picking neighboring pixels for noise reduction in the hybrid selective mean filter (HSMF) method.

The threshold (h) was calculated using the magnitude of the standard deviation (SD) of the pixel values inside an image, which is a measure of noise33. To cover the majority of the image noise in this study, a 3 SD threshold was utilized. An approach proposed in Ref.34 was used to determine the SD automatically. This selects the minimum value of the standard deviation map automatically (SDM) as defined by Eq.(12).

$$SD=mathrm{min}left(SDMright).$$

(12)

The HSMF was supposed to reduce the noise dramatically while maintaining good spatial resolution. The technique is computationally light and fast as it is based on MF, making it easier to employ in clinical imaging than the BF (bilateral filter). Figure9 displays the filtered image by using the HSMF method.

Filtered CT scan images using hybrid selective mean filter method.

To determine the COVID-19, a ML/DL based ensemble classifier was employed35. Four ensemble models are commonly used to create the predictive classifier such as boosting, bagging, stacking, and voting36. The bagging approach of the ensemble methods like a bootstrap aggregation was used in this experiment. To compare the classification performance utilizing the optimized feature vector, three distinct types of classifiers including Long Short-Term Memory (LSTM), ResNet50 and Support Vector Machine (SVM) were employed. These three base classifiers were chosen as they typically outperform other ML/DL techniques. The categorization of any new instance by ensemble approaches is based on the classification votes of the basic classifiers. The output of each base classifier is regarded as a vote, with v=1 for the COVID-19 class and v=0 for the non-COVID-19 class.

The ensemble decision class is one that receives majority of the votes from the base classifiers that means (left(if {sum }_{i=1}^{n}v>frac{n}{2}right)) as indicated in Eq.(13).

$$Ensemble, Class= sumlimits _{i=1}^{n}v,$$

(13)

where the total number of base classifiers is n.

Figure10 represents the ensemble classifier-based bagging approaches where C1, C2, and C3 depict the LSTM, ResNet50, and SVM base classifiers, respectively. Similarly, P1, P2, and P3 signify the votes they represent. The final classification result combines the votes P1, P2, and P3 using Eq.(13) to yield the anticipated class based on the majority votes. To train the base classifiers, the training dataset set was divided into three subsets, D1, D2, and D3, then the testing was performed after training.

The bagging approach in the ensemble classifier.

See the original post:

Ensemble classification of integrated CT scan datasets in detecting ... - Nature.com