Cell Phenotype Classification Based on Joint of Texture Information and Multilayer Feature Extraction in DenseNet

Cell phenotype classification is a critical task in many medical applications, such as protein localization, gene effect identification, and cancer diagnosis in some types. Fluorescence imaging is the most efficient tool to analyze the biological characteristics of cells. So cell phenotype classification in fluorescence microscopy images has received increased attention from scientists in the last decade. The visible structures of cells are usually different in terms of shape, texture, relationship between intensities, etc. In this scope, most of the presented approaches use one type or joint of low-level and high-level features. In this paper, a new approach is proposed based on a combination of low-level and high-level features. An improved version of local quinary patterns is used to extract low-level texture features. Also, an innovative multilayer deep feature extraction method is performed to extract high-level features from DenseNet. In this respect, an output feature map of dense blocks is entered in a separate way to pooling and flatten layers, and finally, feature vectors are concatenated. The performance of the proposed approach is evaluated on the benchmark dataset 2D-HeLa in terms of accuracy. Also, the proposed approach is compared with state-of-the-art methods in terms of classification accuracy. Comparison of results demonstrates higher performance of the proposed approach in comparison with some efficient methods.


Introduction
Detection of cell phenotypes plays an important role in many diferent medical or cell biology applications. Among its most important uses, the following can be mentioned [1]: (i) Analysis of the efects of genes and drugs on screening experiments (ii) Subcellular protein localization (iii) Cancer diagnosis in images acquired using cytological or histological techniques In most related healthcare-systems and medical diagnosis processes, if the type of cells is detected automatically with high accuracy, it can increase the fnal performance of the related system [1]. In the last decade, fuorescence imaging is the most efcient tool to analyze the biological characteristics of cells, which is widely used by specialists [2]. Hence, cell phenotype classifcation in fuorescence microscopy images has received increased attention from scientists in the feld of computer vision and artifcial intelligence. Cells are known to exhibit complex phenotypes such as diferences in shape, texture, and other qualities. Te combination of intricate phenotype diferences together has made cell classifcation and decoding biological processes a complex task. Today, in many laboratories, this task is performed by experienced and specialized persons. Terefore, it is a time-consuming process, and the detection's accuracy is not constant in diferent environmental and time conditions. So a cell classifcation task can be defned as predicting the type of input cells based on predefned cell types in the train set. Te visible structures of cells are usually diferent from each other in terms of shape, texture, relationship between intensities, etc. Terefore, until now, various approaches have been presented for cell phenotype classifcation based on machine learning and computer vision techniques. In this scope, most of the efcient approaches are categorized into two main strategies [1,3]: (i) Handcrafted-based features in joint of machine learning techniques (ii) Deep learning-based methods Cell phenotype classifcation can be categorized as visual pattern classifcation problems. Nearly, in all of the proposed approaches until now, feature extraction and classifcation have been two main phases. Feature extraction is performed before classifcation in machine learning-based methods. However, features are extracted during the learning process in deep-based methods. Also, in some cases, preprocessing is performed to increase fnal performance as a separate phase. In this respect, some approaches use low-level features, such as handcrafted, texture, color, or shape features [4]. Deep neural networks usually extract high-level features during the learning process. So, in some studies, a cell phenotype is classifed based on deep learning models [5].
In the last decade, deep neural networks have grown tremendously and have provided good performance in many computer vision applications. Xu and Qiu [6] have proposed a CNN-based deep neural network for human activity recognition. Tey used the gradient descent algorithm to optimize deep network parameters [6]. Zheng et al. [7] proposed an innovative image denoising method based on the hybrid convolutional neural network with acronym (HDCNN). HDCNN consists of diferent layers, such as the dilated block, Rep VGG block, feature refnement block, and convolution layer [7]. Deep learning-based approaches are widely used in medical image analysis. Shabbir et al. [8] studied many recent studies on glaucoma detection using retinal fundus images. In [8], classical methods based on handcrafted features and deep learning-based methods are studied in terms of methodology and performance. Masood et al. [9] proposed a combination of feature selection flters in joint of diferent supervised classifers for medical data analysis such as hepatitis diagnosis. Yamamoto and Miyata [10] developed a CNN for an automatic cell morphology discrimination technique based on AlexNet. A combination of AlexNet and the transfer learning technique is used in [11]. Te proposed deep neural network comprised three convolution layers, two pooling layers, and two layers for providing the total combination [10]. Deep neural networks are widely used for diferent computer vision applications such as defect detection [12]. Latif et al. [13] presented a comprehensive review on efcient handcrafted features and deep learning approaches.
Classifcation accuracy is the main performance evaluation metric in the cell phenotype classifcation problem.
Terefore, most studies try to increase classifcation accuracy as a fnal aim. Low-level features usually extract the appearance information of the object from the image. Because the appearance of cells is diferent from each other, it is necessary to extract low-level features [4]. For example, two samples of two cell classes (nucleus and actin flaments) are shown in Figure 1. As can be seen, the texture of two samples is fully diferent. Some types of cells are not very diferent in terms of visual appearance, but they have diferences in terms of the relationship between the intensities and texture or shape. In these cases, high-level features can play a vital role in improving fnal performance [5]. For example, the texture and structure of samples shown in Figure 2 are nearly the same.
In this paper, combinations of low-level and high-level features are used in the learning process. First of all, the operator-improved local quinary pattern (ILQP) [14] is used to extract texture features. Next, a multilayer method is presented to extract high-level and deep features using DenseNet [15]. Finally, the normalized combinations of these features are used to feed the classifer. Te main novelty of this paper is to use the combination of texture and deep features to classify the types of cells. An improved local quinary pattern (ILQP) is one of the most efcient operators for image texture description, which was developed in 2020. In this article, for the frst time, ILQP has been used to extract texture features of cells. Also, in this paper, the classifcation layers in DenseNet are removed and replaced by the fatten layer. Te study of published articles shows that for the frst time, the combination of deep features extracted from DenseNet in joint of ILQP features has been used for medical image analysis.
In this respect, the main contribution of this paper can be mentioned as follows: (i) Cell phenotype classifcation based on the combination of low-level texture features and high-level features extracted using DenseNet provides higher performance in comparison with most exciting approaches that use one of the feature types (ii) Extracting deep features from the dense blocks of DenseNet in a multilayer format provide discriminative features for cell phenotype classifcation

Related Works
As mentioned above, the problem of cell phenotype classifcation is grouped as pattern classifcation problems. Terefore, most of the methods that have been presented so far can be categorized into two groups, methods based on machine learning and methods based on deep learning [1]. In most of the machine learning-based methods, feature extraction and classifcation phases are separate from each other, which is due to the nature of these types of classifers.
In the methods presented based on deep learning, the feature extraction process is performed in the learning phase, and the classifcation process is performed in the last layer of the deep network. In the following section, some efective methods in these two felds are examined.

Computational Intelligence and Neuroscience
In a study, Nanni et al. [16] proposed a three-step method for cell image classifcation. In the frst step, rough segmentation is performed to crop background regions. Te authors in [16] suggested the frst step to improve fnal performance. Next, feature extraction is performed in the second step. Tey used basic LBP as a texture analysis operator in this step. Te histogram of LBP values in diferent color channels of the segmented image is quantized to extract numerical feature vectors. In the experiments, Haralick features and TAS feature sets are also performed in joint of LBP. Finally, the classifcation phase is performed in the third step. Tey suggested using the SVM and random subspace neural network (RSNN) as classifers for their experiments.
Liu et al. [17] proposed an approach for cell classifcation in microscopic images based on handcrafted texture features. In this respect, a novel LBP-like operator is presented by assigning an adaptive neighborhood radius in the primary LBP process. Next, a spatial adjacent histogram technique is used to encode microstructures. Some evaluations are performed by the authors in [17] on diferent medical datasets such as the HeLa cell phenotype dataset. Te reported results show higher performance for the proposed operator than for the primary LBP version and some other texture-based operators.
Fekri-Ershad [18] proposed a new LBP-like operator called multithreshold uniform local ternary patterns (MT-ULTPs) to extract discriminative texture features. As reported in [18], the noise sensitivity of MT-ULTP is lower than that of many other LBP-like operators, such as LTP [19], because of using diferent binarization thresholds in the local pattern building process. Also, it can be performed in diferent radius sizes to extract micro-and macro-local patterns. MT-ULTP in joint of the random forest as a classifer provides an 86.77 percent accuracy for cell phenotype classifcation for the 2D-HeLa dataset [18].
An ensemble method of diferent CNNs in terms of depth and structure is used in combination with a feature concatenation technique [20] for medical image classifcation. Also, a PCA technique (with 95 percent) and transfer learning are performed to improve the fnal performance of the proposed ensemble method. Te authors evaluated the performance of their proposed approach on the Pap smear and cell phenotype images. Results show the classifcation accuracy about 93.51 percent for the 2D-HeLa cell dataset [20]. Te standard deviation of the classifcation accuracy was 2.29 percent, which is not suitable for medical diagnosis problems.
Zhang and Zhao [21] used the CapsNet neural network for classifcation of ten phenotypes of cells. Tey used 3 capsules in the structure of CapsNet which performed convolution and reshape and lambda layers in a sequence format. Capsules in CapsNet were trained to capture the possibility of certain features and variants rather than to Computational Intelligence and Neuroscience capture the characteristics of a specifc variant. Te redesigned CapsNet in [21] provided a classifcation accuracy of 93.08 for the 2D-HeLa cell phenotype dataset. Nguyen et al. [22] presented an innovative deep network architecture using transfer learning for microscopic image classifcation such as cells. Tey concatenated total features that are extracted using 3 diferent pretrained CNNs in the learning phase. Finally, two fully connected layers are fed using these features to perform the classifcation phase. Inception-V3, ResNet152, and Inception-ResNet-V2, are three pretrained CNNs that are used in the learning phase [22]. High complexity and respected higher runtime, because of using three CNNs, are two main limitations of the proposed method [22]. Also, the number of parameters of this method is too high. So pretrained weights may not provide high accuracy in other related applications.

Proposed Cell Phenotype Classification Approach
As mentioned above, cell phenotype classifcation can be categorized as visual pattern classifcation. Terefore, our proposed approach included two main phases, feature extraction and classifcation. Te block diagram of our proposed approach is shown in Figure 3. Each box is described with details in the following sections.

Improved Local Quinary Patterns.
Texture is one of the main properties which humans use to defne images. Te image texture provides efcient features for content, objects, background, contrast, internal relationship, etc [23]. As shown in Figure 1, diferent types of cells difer from each other in terms of their visual texture. Terefore, extracting texture features may play an important role in cell classifcation. Until now, various methods have been presented for image texture analysis. Most of the proposed operators so far have described the image texture as a set of numerical features. Te method presented in this paper does not depend on a specifc type of texture analysis operator. Terefore, most of the operators that can extract numerical features can be performed. Te local binary pattern (LBP) is one of the most effcient descriptors to extract texture features. Te initial version of this operator was presented in 2000 by Ojala et al. [24]. In the last two decades, many diferent modifed versions of this operator have been presented, each trying to cover each other's weaknesses. One of the most recent LBP-like operators that provided acceptable performance in texture analysis is an improved local quinary pattern (ILQP) [14]. Te initial version of LBP produces binary patterns in each local neighborhood. Nanni et al. [25] suggested a quinary pattern to reduce the noise sensitivity of the initial LBP version called the local quinary pattern (LQP). An initial LBP operator uses the central point intensity as a fxed threshold to produce binary patterns: Terefore, even weak noise, if it is placed on the image, leads to a change in the intensity of pixels (including neighborhood centers), and the efciency of the operator drops drastically. LQP uses two fxed thresholds (T 1 and T 2 ) to create a quinary pattern with 5 digits [25]. It is performed using a new transform function as follows: In order to reduce complexity, Nanni et al. [25] proposed a transform function B c (x) to convert LQP into four binary patterns with acronym B 2 , B 1 , B −2 , and B −1 : Similar to LBP, each binary pattern is mapped to decimal numbers. Terefore, four decimal numbers can be produced for each neighborhood. Finally, four histograms can be computed for each image, which can be used as a texture feature set. An example of converting the quinary pattern to four binary patterns is shown in Figure 4.
An improved local quinary pattern (ILQP) is proposed in [14] to reduce limitations of LQP operators. As mentioned above, thresholds should be selected by users in the LQP process. A dynamic algorithm is described to defne LQP's thresholds in ILQP [14]. To choose the frst threshold (T 1 ), the global median absolute deviation (GMAD) is suggested for ILQP. GMAD is a statistical criterion that measures how a set of data (such as intensities) is developed. GMAD is calculated as the average of local MADs over the whole image, so it is not sensitive to noise. If noise causes changes in intensities of some pixels, the efect of disturbance on the average of the total data (all pixels) will be very low. GMAD can be calculated as follows: where P is the total number of neighbors in each local neighborhood and g i shows the intensity value of the i th neighbor. Also, N is the total number of possible neighborhoods in the image based on the radius of the ILQP operator. Te global signifcant value (GSV) was proposed for the frst time by the authors in [14] for medical image analysis. However, in some other studies, it is used as a discriminative statistical feature in learning phases. It is suggested that GSV should be considered as the second threshold (T 2 ) in the ILQP process. GSV can be computed based on the local signifcant value (LSV). LSV calculates the diference between intensity of neighboring centers and surrounding neighbors. Te total average of LSV for the whole images is called GSV. It can be calculated as follows: where M × N shows the image size. Also, LSV i, j means the local signifcant value of the neighborhood with a center pixel c with coordination (i, j). Computational Intelligence and Neuroscience

Proposed Multilayer Feature Extraction in DenseNet.
DenseNet was proposed in the CVPR conference (2017) as an innovative CNN-based structure [26]. DenseNet, instead of most CNN-based networks such as ResNet, CIFAR, ImageNet, and AlexNet, utilizes dense connections between internal layers using dense blocks. DenseNet contains shorter connections between its layers close to the input and those close to the output. Terefore, in most cases, DenseNet can be more accurate in the training process. ResNet [27] uses skip connections that bypass the nonlinear transformation, but DenseNet adds a direct connection from any layer to any subsequent layer. So the i th internal layer receives the feature maps of all former layers. Terefore, all layers can be connected (with matching feature map sizes) directly to other ones. In recent works, DenseNet has provided acceptable results for classifcation problems such as medical diagnosis. Te main reason for the success of this network is that it is deeper than some common networks and its optimization is easier. DenseNets have some advantages such as  Table 1. As can be seen, the main diference is in the number of internal layers of each dense block. Also, feature maps are highlighted in Table 1 in terms of output size.
Te basic DenseNet framework contains four dense blocks, with each block comprising four convolution layers. Furthermore, each convolution layer applied four techniques batch normalization (BN), ReLU activation, squeeze, and excite operations. According to the initial flter size and stride value, after each step, the feature vector size increases due to the concatenation operation. Te transition layer was introduced after each DB to perform the downsampling step to solve this problem. Te N th layer of the framework has N inputs as the N th layer takes the outputs of all previous N − 1 layers: where [I 0 , I 1 , I 2 ,. . ., I N − 1 ] are the feature maps from the previous N − 1 layers, which are connected to the N th layer and indicated by I N . Furthermore, the transition layer comprised convolution and pool layers. Te bottleneck layer contained a 1 × 1 convolution layer, which was employed to minimize the size of feature maps and enhance computational efciency. In most cases, the fnal feature map of DenseNet is used to classify the input image. In this study, we suggested using multilayer feature maps in a concatenating format for the classifcation phase. Te output feature map of each dense block is frst entered into a new average pooling layer, and after dimension reduction, the output is entered into a fatten layer. Te fattening process is commonly used to convert all the resultant two-dimensional feature maps from pooling/convolutional layers into a single long continuous linear vector. Te fattened matrix is fed as input to the fully connected layer to classify the images in CNNs. As a result, the feature map in the matrix format is converted to a feature vector in low dimensions. A visual example of the fatten layer is shown in Figure 5. Te primary layers in all CNN-based networks usually show low-level and low-depth properties of the image. In our proposed approach, low-level features are extracted using the ILQP operator. Terefore, to reduce the dimensionality and computational complexity, the output of the frst dense block is not included in the suggested feature extraction process. Te structure of our proposed feature extraction approach is shown in Figure 6. Te main innovation of this paper is to combine texture features and deep high-level features to classify cell types. Figure 3 shows the train and test phase of the proposed method with focus on innovation. As described in this section, to extract deep high-level features, the structure of the classifcation layers in DenseNet has been removed and new layers have been added. In Figure 6, the feature extraction phase and how to feed the classifer are shown with focus on this point. As described, classifcation layers of DenseNet have been removed in the proposed approach. Terefore, we did not use them directly for classifcation. In other words, DenseNet is used to extract deep high-level features in our proposed approach. Feature maps are not very diferent in DenseNet versions that do not difer much in terms of the number of layers. Terefore, four versions of DenseNet (D.Net-121, D.Net-169, D.Net-201, and D.Net-264) have been selected in such a way that the number of internal layers that play a role in making the feature map increases from low to high.
As described above, four diferent DenseNet versions are evaluated in this paper, which are DenseNet-121, DenseNet-169, DenseNet-201, and DenseNet-264. All the above networks have been trained by Adam optimizer in 20 epochs. Te focal loss function with c = 2 is used for the training process [28,29]. Te initial learning rate is considered to be 10 −5 for frst 10 epochs. Ten, it decreases to 10 −6 for next 10 epochs. We select the best loss function, which is a hyperparameter, which depends on the problem we are facing. Multiclass cross-entropy loss seems a wise choice for many studies. However, according to the amount of distribution of cells in the body of each human and according to the database samples, we are facing an imbalanced classifcation problem. Focal loss is another choice that we can leverage its properties to enhance the performance of our model. Tis loss function tries to generate the classweighting system in order to balance the samples in each batch size of data (Equation (11)): Here, pt is a function of true labels. Focal loss can be interpreted as a binary cross-entropy function multiplied by 6 Computational Intelligence and Neuroscience  Figure 6: Te block diagram of the proposed approach.
Classifcation layer 1 × 1 7 × 7 global average pool 1000D fully connected, softmax Computational Intelligence and Neuroscience a modulating factor (1 − pt) c which reduces the contribution of easy-to-classify samples. Weighting factors α t balance the modulating factor. Tere are several approaches for incorporating focal loss into a multiclass classifer. One-versusthe-rest (OvR) technique is used in this paper, in which a binary classifer is trained for each class C. Te data from class C are treated as positive, and all other data are treated as negative.

Dataset.
To evaluate the performance, experiments were carried out on benchmark cell phenotype datasets called 2D-HeLa [30] and human epithelial type-2 (Hep2). Te 2D-Hela database contains 862 images that are categorized into 10 classes of cell types. Te size of all images is the same and equal to 382 × 382. All images are saved in the TIFF format. Te size of classes is not balanced, and there are between 73 and 98 samples in each class. Te vertical and horizontal resolutions of all images are 72dpi and scale variation can be seen in images. Some examples of the 2D-HeLa dataset are shown in Figure 7.
Te human epithelial cell type 2 (Hep2) dataset contains 63,445 cell images in six classes. Tis dataset is collected by single-cell segments from 948 images from a set of multicell images submitted at the ICPR 2014 HEp-2 cell competition. Te classifcation task in this paper is to distinguish diferent staining patterns in Hep2 images with indirect fuorescence to indicate antibodies associated with autoimmune diseases. Te six classes in this dataset are homogeneous (HO), speckled(SP), nucleolar(Nuc), centromere(CE), Golgi(GO), and nuclear membrane(Num). Some examples of Hep2 dataset samples are shown in Figure 8.

Performance Evaluation Metrics.
As mentioned above, cell phenotype classifcation can be categorized as visual pattern classifcation problems. Terefore, accuracy is the main metric to evaluate the performance of related studies. Cell classifcation is a multiclass problem. Hence, the classifcation accuracy can be calculated using the confusion matrix as the sum of correct cells in the table (the number of correct label-predicted test samples) divided by all cells in the table (total number of test set samples). In other words, accuracy is the most intuitive performance measure, and it is simply a ratio of correctly predicted observation to the total observations. Also, the performance of our proposed approach is evaluated in terms of precision and F1 score. All of the performance-evaluated measures are described in Equations (12)- (15), where TP means true positive, FP means false positive, and FN shows false negative. Cell phenotype classifcation is a multiclass classifcation problem, so measures are evaluated as an average of all classes: accuracy � correctly predicted observation total observation × 100, (12) precision

Performance Evaluation of the Proposed Approach.
Multilayer feature extraction in DenseNet is proposed in this paper. Our proposed approach can be performed in each type of DenseNets. Only the number of extracted features may be diferent. In this respect, we evaluated our proposed approach based on diferent DenseNet types in terms of accuracy. Te combination of texture and deep features is extracted in this paper. Terefore, most of the numericalbased classifers can be used in the classifcation phase. We evaluated our proposed approach using diferent efcient classifers as mentioned in Tables 2 and 3.
As can be seen in Tables 2 and 3, DenseNet169 provides the highest accuracy compared with other versions. As mentioned above, the Adam optimizer with the focal loss function is used in this paper to optimize hyperparameters. Te efectiveness of the proposed method is evaluated based on diferent learning rates, and fnally, a learning rate of 10 −5 in the frst 10 epochs and a learning rate of 10 −6 in the next 10 epochs provided the highest classifcation accuracy. In Table 4, the evaluation results of the proposed method are presented based on DenseNet169 in terms of diferent learning rates. [31].

Comparison with State-of-the-Art Methods.
Te performance of the proposed method is compared with some efcient approaches in terms of accuracy as reported in Tables 5 and 6. In this respect, to have a fair comparison, the same validation conditions (K-folds) and same dataset (2D-HeLa and Hep2) are considered for comparison experiments. In order to have a fair comparison, all of the compared results in Tables 5 and 6 are reported based on results in related papers. In some studies, standard deviation is not reported. Efcient and up-to-date methods have been chosen from both common strategies (handcrafted and deep-based). Te comparison results show that our proposed method provides a higher accuracy rate than other state-of-the-art methods in the related scope. Te ensemble method of inception-v3, ResNet152 and inception-ResNet-v2 [20] provides about 0.15 percent higher accuracy than our proposed approach. As can be seen, the standard deviation of our proposed approach is less than about 0.44 percent [20], which shows that our proposed approach is more stable in experiments. Also, the proposed approach in [20] is an ensemble method of three diferent CNNs, which increase the runtime and computational complexity. In this paper, a combination of handcrafted texture features and deep features extracted from modifed DenseNet has been used to classify cell types. Cell-type classifcation is performed on the same database using three types of pretrained DenseNet. Te related results are shown in Table 5 as the baseline. Te   8 Computational Intelligence and Neuroscience   Computational Intelligence and Neuroscience  [18] Handcrafted 86.77 --Haralick-based SVM [21] Handcrafted 84.10 --Random subspace of LMC [32] Handcrafted 90.24 --AdaBoost ERC [32] Handcrafted 91.53 ± 0.02 --SIFT (BoW (LLC) + SPM + softmax) [33] Handcrafted 89.37 ± 1.5 --LTP 16,2 [34] Handcrafted 87.00 --LBP-rotation invariant uniformLBP 16,2 [34] Handcrafted 82.70 --Orthogonal locality preserving projection (OLPP) [34] Handcrafted 89.30 --Neighborhood preserving embedding (NPE) [34] Handcrafted 93.20 --Discriminative LBP [34] Handcrafted 84.5 --LBP-rotation invariant [34] Handcrafted 75.01 --Completed LBP [34] Handcrafted 88.8 --Random subspace ensemble of Levenberg-Marquardt neural network (RSNN) [16] Classic neural nets 85.00 --   10 Computational Intelligence and Neuroscience results show that the proposed combinational method provides higher accuracy than baseline DenseNet. Also, the performance of the proposed approach is evaluated in terms of precision and F1 score. Some of the compared methods reported their performance in terms of F1 score as can be seen in Tables 5 and 6.

Conclusion
Te main goal of this study was to propose an efcient approach for cell phenotype classifcation. Unlike most popular approaches, low-level and high-level features are performed in joint of format in the learning phase. ILQP is performed to extract handcrafted texture features. In order to extract high-level features, an innovative multilayer feature extraction method based on DenseNet is proposed. Te proposed approach provides about 95.59 percent classifcation accuracy for the Hep2 dataset and 93.36 percent accuracy for the 2D-HeLa dataset. Te experimental results show the higher performance of the proposed approach compared with that of state-of-the-art methods in this literature for HeLa and Hep2 datasets. Also, the results prove that the combination of texture and deep features has higher accuracy than using the baseline deep network alone, with more than 0.7 percent for benchmark cell datasets. Te runtime of the proposed approach is lower than that of popular CNN-based methods because of removing fully connected and softmax layers from the end of the deep structure. One of the important advantages of the presented method is its generalizability to all kinds of feature extraction operators. Te proposed method for combining texture and deep features has a general structure. It means that instead of the ILQP operator, any other operator that extracts features in vector form can be used to combine with deep information. Te method presented in this paper has two phases of learning and classifcation. Terefore, as a suggestion for future works, it can be used for many diferent problems of visual pattern classifcation for the computer vision scope.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that they have no conficts of interest.