Computer Tomography Image Based Interconnected Antecedence Clustering Model Using Deep Convolution Neural Network for Prediction of COVID-19

Residents of Wuhan in Hubei Province, China, have been struck down by COVID-19, an unusual coronavirus infection discovered by the World Health Organization (WHO). According to prior studies, the virus may have originated in a fish market in Wuhan, and bats may have served as the vehicle for its transmission from bats to humans [1]. The World Health Organization has classified this severe and fast-spreading virus from the Coronavirus family, an emergency because of the lack of health facilities and resources [2]. COVID-19, which originated from the SARS-CoV-2 virus, has posed a serious and immediate threat to human health. Since its first appearance in China's Hubei province in early December, the flu virus has quickly spread around the world [3]. With more than 170 nations reporting cases as of December, 20, 2021, it is probable that the true number of infected people is far higher. Nearly 3,457,000 people have been died because of COVID-19 [4].

There has been a significant shortage of medical supplies and an increase in the demand for hospital beds as a result of this pandemic. Therefore, it is essential to make rapid clinical decisions and make effective use of healthcare resources. COVID-19 RT-PCR [5], which has been in short availability in poorer countries for a long time, is the most extensively utilized diagnostic test [6]. Because of this, preventative actions are delayed, which raises the probability of infection. The effects on healthcare systems can be reduced if COVID-19 is diagnosed quickly and accurately by screening [7].

In the face of dwindling healthcare resources, medical professionals and organizations are relying on these models to help them prioritize their patients' care in a more informed manner [8]. CT scans, for example, are included in some of these models, as is the integration of these aspects. Because previous models relied on inaccurate data from inpatients [9], it is difficult to conduct SARS-CoV-2 screenings in the general public [10]. As may be seen in Figure 1, a high-resolution CT scan of a lung is represented.

Deep learning algorithms can tell the difference between a healthy individual and a COVID-19 patient based on a CT scan of the suspected patient. COVID-19 diagnostic systems are being built using deep learning algorithms. DenseNet121 [11], VGG16 and EfficientNet are just a few of the algorithms that use multiclass classification [12]. COVID-19-positive patients as well as ordinary patients and others are considered. CT scans indicate disease of the lungs, such as pneumonia and influenza, which should be considered a distinct disease. COVID-19's detection inaccuracy and lack of testing kits need the development of other testing procedures, including x-rays and chest CT scans [13]. Using these tools, it is feasible to identify the disease's radiographic characteristics [14]. Hospitals typically have x-ray machines; therefore, radiologists prioritize using them. When it comes to the chest, X-ray scans can't tell the difference between hard and soft tissues [15]. It is for this reason that chest CT scan images are used since they can detect soft tissues with more accuracy and produce results with greater speed Using deep learning techniques [16], researchers have been able to automatically analyze CT scan images and determine if a person is COVID positive or not [17].

Two algorithms that make extensive use of lung imaging in their diagnostic processes are the convolutional neural network (CNN) and the deep neural network (DNN). The CT scan image is the first step in corona virus detection [18]. A number of processes will be applied to the image to help identify the issue. There are three steps involved in feature-based classification: pre-processing, morphological building [19], and image classification. Instead of using sophisticated data collection strategies, features are manually gathered in an experiment and supplied into a classifier. The most common types of ensemble learning methods are banded, boosted, and layered [20].

Ensemble learning has the potential to significantly improve the generalization of a learning system. In order to develop classification model, the most common approach is to apply multiple learning algorithms to the same dataset. Distinct classifiers can be constructed using this method. Using the same supervised learning on different training data is also a possibility [21]. As a basic classifier, this approach produces a homogeneous classifier. One of the most important aspects of classification is the integration of learning techniques and averages into the overall strategy. Based on the target use of integrated learning, a variety of combination tactics are frequently employed. Individual learners' predictions are averaged or weighted when integrated learning is used to estimate regression [22]. Using integrated learning approaches, the classification findings must be voted on individually in order to arrive at a final classification. Absolute and relative majorities are the two sorts of majorities that can be voted on [23]. As a result of the overall majority voting technique, the final classification of integrated learning output is selected by more over half of the individual learners [24]. The relative majority technique, used to determine the largest number of unique learners who provide a specific classification result, is used to classify an integrated learning output. Figure 2 depicts the COVID prediction process utilizing the ensemble technique with hidden layers.

1.png

Figure 1. Lung CT images

2.png

Figure 2. COVID prediction process

COVID-19 disease, also known as COVID-19, is detected using chest CT images and a novel deep neural network method. Here, a convolutional neural network (CNN) was used to analyze CT scans of the lungs. This is accomplished with the help of ensemble learning and the DenseNet201 CNN architecture. The design's effectiveness in classifying CT images requires the use of multiple activation strategies [25]. To reduce diagnostic time, radiologists are increasingly turning to deep learning models trained on CT images. Radiologists were able to better distinguish COVID-19 abnormalities from chest CT pictures with the help of CNN designs, prompting the initial proposal of natural image analysis [26]. Regular CT scans and a huge dataset of illnesses not caused by the COVID virus or bacterium [27] were also examined by the researchers alongside the COVID-19 X-ray images. For precise COVID diagnosis, the suggested technique employs a deep learning model, which aids medical practitioners in administering timely, effective treatment.

When trained on a small sample of photos, a simple CNN design can outperform more complicated networks like Xception and DenseNet. Even though the classification accuracy of CNNs is high, therapists should hold off on drawing any conclusions about the input image until they can examine the region that the CNN identified as being relevant. In this research, an Interconnected Antecedence Clustering Model employing DCNN model is proposed that accurately extracts the features for prediction of COVID-19. The proposed model considers image dataset and performs segmentation on the images. The bagging ensemble model based prediction set is generated for the accurate prediction of COVID.

In order to detect problems in the lungs, non - linearities in the lung can indeed be identified via feature extraction. Many feature extraction algorithms have been activated in order to better detect COVID-19. More often used methods for customized feature extraction include the Discrete Wavelet Transform (DWT), Gray Level Co-Occurrence Matrix (GLCM), and Haralick texture features. CNN-based feature models have also been taken into consideration in the extraction process.

By keeping only the most relevant data, feature selection reduces the amount of data that isn't strictly necessary. Several algorithms have been developed to prune irrelevant data from feature representations. The Stochastic Gradient Descent (SGD) model could be used with some fiddling with the learning rate, speed, and batch size. A 512×512 image was delivered to the model after the RGB channels were rearranged. We scaled the images to 512×512 for the training set, then randomly split them into 480×480 squares, and standardized the results. The images were then normalized after being flipped horizontally. Some of the features extracted from the images are id, diagnosis, radius_mean, texture_meanperimeter_mean, area_mean, smoothness_mean, compactness_mean, concavity_mean, concave points_mean, symmetry_mean, fractal_dimension_mean, radius_se, texture_se, perimeter_se, area_se, smoothness_se, compactness_se, concavity_se, concave points_se, symmetry_se, fractal_dimension_se, radius_worst, texture_worst, perimeter_worst, area_worst, smoothness_worst, compactness_worst, concavity_worstconcavepoints_worst, symmetry_worst, fractal_dimension_worst.

Using a majority voting-based ensemble learning technique, the final class predicted by either model is employed in this study. The inceptionV4 model is capable of factoring both convolutional and multiple-size filters at once. It performs a 1x1 convolution before each of these filters in order to limit the number of input channels. As a result of scaling the neurons in each layer consistently, the proposed model results in a more uniform scaling of the network's depth and width. The proposed model work flow is shown in Figure 3.

3.png

Figure 3. Proposed model workflow

In this analysis, Bagging is employed, shorthand for the bootstrap aggregating ensemble. This deep learning ensemble approach was created to improve classification accuracy and consistency. When compared to the Boosting ensemble method, bagging helps avoid overfitting problems because the Boosting algorithm only uses the misclassified samples from the previous phase as training data. Using the Bagging ensemble technique, we fine-tune the mode to predict the class probabilities of observations in the test set, and then we use these refined models to compute an average probability score. In this research, an efficient CT Image based Interconnected Antecedence Clustering Model using DCNN (IACM-DCNN) is proposed for accurate classification and prediction of COVID-19 from CT images. The process of COVID detection is clearly explained in the algorithm.

Algorithm IACM-DCNN

{

Input: CT Image Dataset{CTIDS}

Output: Prediction Set

Step-1: Load the image from the dataset and perform image segmentation to divide the image into partitions for accurate extraction of features from the image. The image segmentation process is performed as

$\begin{aligned} \text { contrast } & =\sum_{p, q=0}^N \operatorname{CTIDS}(p)_{p, q}(\operatorname{MinP}-\operatorname{Maxq}+\operatorname{greylevel}(\operatorname{img}(p))\end{aligned}$

$\begin{gathered}\text { Dissimilarity }=\sum_{\substack{p, q=0 \\}}^N \operatorname{img}|p-q|+\operatorname{MinP}(p, q) -\operatorname{sim}(q, p) \end{gathered}$

$\begin{array}{r}\text { Entropy }=\sum_{p, q=0}^N-\ln \left(\operatorname{Min} P_{p q}\right) * \operatorname{Min} Q_{p q}+\max (\operatorname{contrast}(q, p))\end{array}$

$\operatorname{SegSet}(\operatorname{CTIDS}(P, Q))=$

$\sum_{P, Q=0}^M \frac{\operatorname{Min} P(p, q)+\sum_{p, q=0}^{M-1}\, \,\,\,\operatorname{MaxQ}_\,\,{p \,\,q}\,\,\,(q-p)^2}{(q-p)^2+T} + \sum_{p, q} \frac{\mid \text { maxgreylevel }(\mathrm{q}, \mathrm{p}) \mid}{|\operatorname{sizeof}(C T I D S)|}\,\,\,\,+\max (entropy)$

Here $\operatorname{Min} P, \operatorname{Max} P$ are the minimum and maximum intensity levels of an image, p and q are the image pixel coordinates.

Step-2: The segments are considered for feature extraction and the features are extracted from the image that are used for training the ensemble bagging model. The feature extraction is performed as

$\begin{aligned} & \operatorname{Attr}(\operatorname{SegSet}(p))= \Sigma_{p=1} \frac{\text { CTIDS(img(i) ) }+ \text { minrange (Segset }(\mathrm{p}, \mathrm{p}+1)}{\text { maxintensity }(\operatorname{SegSet}(\mathrm{p}))}+ \max (\operatorname{contrast}(p, p+1)) \\ & \end{aligned}$

$\begin{aligned} & \text { FeatVec }[L] =\sum_{\substack{\text { img } \in \text { CTIDS }(p) }} \frac{\operatorname{sim}(\operatorname{Attr}(p, p+7)-\text { minintensity }(\operatorname{SegSet}(p, p+7)))}{\operatorname{sizeof}(\operatorname{Attr})}\,\,\,\,-\min (\text { entropy) }\end{aligned}$

Step-3: The clustering model is applied on the extracted features so that similar values are grouped together that are used for processing and disease detection. The clustering process is performed as

$\begin{aligned} & \text { IS }(\text { FeatVec }[L])= \\ & \frac{\sum_{p=1}^N(\operatorname{sim}(\operatorname{Attr}(p, p+1))+\operatorname{diff}(\text { FeatVec }(\min , \max ))-\min (\text { int ensity })}{\sqrt{\sum_{p=1}^N \max (\text { FeatVec }(\operatorname{SegSet}(p)))+\operatorname{sizeof}(\text { FeatVec })}} \\ & + \text { range }(\operatorname{sim}(\text { Attr }))\end{aligned}$

$\begin{aligned} \operatorname{Clset}(I S(L))= & \operatorname{diffrange}(\operatorname{IS}(p, q))+\sum_{p=1}^N \operatorname{diffAttr}(\operatorname{maxFeatVec}(p, p+1))+\operatorname{unique}(\operatorname{FeatVec}(p+1, p+2))\end{aligned}$

Step-4: The Bagging ensemble model is considered that works based on the voting model used to train the model by considering the 3×3 convolution hidden layers. The layers consideration is performed as

$\begin{array}{rl}H L(L, C)=\sum_{p=1}^N & C l \operatorname{set}(\operatorname{getAttr}(p, p+1)) -\operatorname{diffAttr}(p+1, p+2)+\text { FeatVec }(\operatorname{sim}(p, q))))+\frac{3 * \sqrt{\operatorname{sizeof}(C T I D S[L])^{q-p}}}{\operatorname{sizeof}(\text { Clset })}\end{array}$

Step-5: The feature class specification as COVID and non-COVID is performed by the last 1×1 convolution layer considered as output in the class labelling process. The class specification base model training and cross validation for function approximation is performed as

$\begin{aligned}\operatorname{Pset}(\operatorname{Clset}(L))=\sum_{p=0}^N \max (\operatorname{FeatVec}(q, p))+\frac{(\operatorname{Max}(\text { FeatVec }((\operatorname{sim}(q, p))))-\operatorname{Min}(\operatorname{FeatVec}(\operatorname{sim}(q, p)))}{\left(\frac{N+\lambda^* p}{3}\right)\left(\frac{3^* C-\lambda^* q}{3}\right)}+\,\,\operatorname{Th}\end{aligned}$

Here Th is the threshold value considered for feature class specification calculation process. λ is the max intensity value. The features that are dissimilar to the original values to predict the COVID is evaluated.

Step-6: The meta training model that reduces the cross functional approximations are trained with the base model training reducing the false predictions is performed as

$\begin{aligned} & \text { baseModel }(\text { Pset }(L))=\left(\frac{\lambda}{2^*} \text { Maxrange }(\text { FeatVec })\right)^{\frac{1}{\lambda}}+\left[\sum_{p=1}^N\left(\frac{1}{(\lambda-\max \operatorname{Attr}(\operatorname{Pset}(p)))}\right)^{\lambda / 3}\right]\end{aligned}$

MetaModel(baseModel(C))

$\begin{aligned} & \Rightarrow \frac{N}{\lambda^2}- \left(\sum_{p, q, i=1}^N \min \operatorname{Attr}(\text { baseModel }(L(p, p+1)))+\left(\exp \left(\frac{\left\|p_i-q_i\right\|^2}{\lambda^2}\right)\right)\right) \end{aligned}$

Step-7: The predictions based on the bagging ensemble model with the meta training is generated as

$\begin{aligned} & \text { Pr edSet } \\ & =\frac{\sum_{i=1}^N \operatorname{Min}\left(Clu_ \_\operatorname{Set}(i)\right)+\operatorname{sim}(\max (\text { weight }(i), \text { weight }(i+1))}{\max (\operatorname{Weight}(i))}\end{aligned}$

Step-8: Display the prediction set.

}