Se-Resnet: A Novel Method for Gastrointestinal (GI) Diseases Classification from Wireless Capsule Endoscopy (WCE) Images

For the categorization of biomedical pictures, DL and ML techniques have been widely utilized. These classification techniques support medical professionals in making precise diagnoses and treatment recommendations. The methods currently used to categorize GI illnesses are reviewed in this section (Table 1). An AI approach is suggested by Öztürk and Özkaya [16] for classifying GI tract picture datasets with a limited amount of labeled information. The foundation of the suggested AI technique is the CNN structure, which is acknowledged as the most effective automatic categorization method currently in use. This method assumes that to categorize unbalanced datasets robustly, a shallowly trained CNN structure needs to be backed by a powerful classifier. The features from every pooling layer in the CNN structure are sent to an LSTM layer for this reason. By integrating all LSTM layers, a categorization is created.

Sharif et al. [17] suggested a method based on the merging of geometric and deep CNN features. Initially, a method called contrast-enhanced color features is utilized to extract disease zones from provided WCE images. Segmented disease regions are used to extract geometric features. Then, based on the Euclidean Fisher Vector, a special VGG19, and VGG16 deep CNN feature fusion is carried out. Unique features and geometric features are combined, and the best features are then chosen using a conditional entropy technique. K-Nearest Neighbor is employed to categorize the final features that have been chosen.

Using endoscopic images, Mohapatra et al. [18] suggested a hybrid EWT and CNN technique for GI illness identification. The primary goal is to identify abnormal GI tract disorders, specifically Ulcerative Colitis, Barrett's, Polyps, Esophagitis, and Hemorrhoids. As a result, only abnormal and normal pictures from the dataset are considered. The initial step includes some procedures, including image scaling, picture augmentation, and picture pre-processing, which aid in preparing the dataset and resolving some issues before other processing. The use of EWT, which helps to break down the pictures into their IMF, which serves as a picture feature pattern extractor is the next stage. The deep network is trained using these extracted patterns. The system is tested and trained on two different levels. Classifying the abnormal and normal pictures is the aim of the first stage of classification. Deliberately removing the normal pictures, aids in the subsequent phase's concentration on the diseased class pictures. Only the aberrant pictures from level one's second level are further categorized into the previously mentioned five distinct illness classifications.

Table 1. The merits and demerits of previous studies on GI

To categorize various gastrointestinal disorders, Ramamurthy et al. [19] suggested an automated categorization method based on DL. The suggested models are trained using the HyperKvasir tagged pictures dataset. To increase the number of samples for improved generalization, the input images are initially enhanced. Two separate networks, EfficientNet B0 and the suggested Effimix network received these modified samples as input. Dropout regularization and feature concatenation were both used after the features from these two networks were joined. The input gastrointestinal picture sets are divided into 23 classes by the suggested approach.

To create a model for the diagnosis of gastrointestinal diseases, Haile et al. [20] suggested concatenating the retrieved attributes of the InceptionNet and VGGNet networks. To extract features from the provided endoscopic pictures, the DCNN’s InceptionNet and VGGNet are trained and deployed. Then, using machine-learning classification methods, these retrieved features are concatenated (KNN, SVM, Random Forest, and Softmax). Utilizing the provided standard dataset, SVM is one of these approaches, outperforming the others in terms of performance.

1.png

Figure 1. The proposed methodology architecture

Finally, to classify the eight classes of gastrointestinal diseases, utilize a deep learning technique of SE-ResNet with Bald Eagle Search optimization technique to get better accuracy of results. Figure 1 illustrated the architecture of the proposed technique.

4.1 Pre-processing

During the pre-processing stage, reduce the noise from the pictures. We remove noise from the input pictures to enhance the diagnosis of GI disease. Here, mean filtering is used to initially reduce the noise in the input pictures. Using numerous picture smoothing templates for picture convolution processing is a standard method for picture de-noising in the spatial domain to reduce or reduce noise. The fundamental idea behind mean filtering is to replace a pixel's single grey value with the total of the grey values of all of the pixels around it. The picture is considered as g(x, y) after smoothing and mean filtering, and it is determined for a pixel point (x, y) in a given picture with f (x, y), where its neighborhood S contains M pixels, by the following Eq. (1):

$G(x, y)=\frac{1}{M} \sum_{(i, j \in S)} f(x, y)(x, y) \notin S$             (1)

4.2 Feature extraction

After noise reduction is completed, we use the DenseNet-121 method to extract the features. This research proposes to utilize the DenseNet-121 approach to extract features and categorize GI diseases. Convolution and pooling layers from the DenseNet-121 method are used to extract features like shape and position from input wireless capsule endoscopy images.

Each layer is linked to every layer below it by a CNN named DenseNet-121. A Dense Block in CNNs is a module that connects all layers directly. Use a dense connection strategy to achieve feature extraction as the input for the subsequent layer, where every layer is connected layer and to each layer before it on the channel dimension [22]. A dense convolution neural network of the type DenseNet-121 is the one being used in this situation. The DenseNet design objective is to make DL networks significantly more profound while simultaneously making them simpler to develop by employing fewer interactions between the layers. Every layer in the DenseNet CNN is connected to any non-subsequent levels that are deeper in the hierarchy. This is carried out to facilitate the greatest data stream between organizational levels. To maintain the feedforward nature of the system, each layer compiles inputs from all preceding layers and offers its component direction to all layers that will follow it.

As a result, all of the initial convolutional blocks in the "$i^{t h}$" layer has highlight guides and "I" information sources. The element maps for all succeeding "I-I" levels are provided. This presents an alternative to the simple "I" links found in typical profound learning models relationships within the company and it represent in Eq. (2).

$(I *(I+1)) / 2$                  (2)

Since it is unnecessary to learn trivial element maps, it employs fewer boundaries than traditional convolutional neural networks.

DenseNet features two vital elements in addition to the essential pooling and convolutional layers. The levels are called Transition Blocks and Dense Blocks. When the amount of filters varies between Dense Blocks, the elements of the channel are enlarged. The development rate (k) helps to generalize the $l^{t h}$ layer. Each layer's data addition is controlled by this. In a nutshell, it is the amount of channels that a dense block outputs. This suggests that a dense layer's (l's) share of features with the dense layer (l-1's) preceding it is k[l]. This is referred to as the growth rate because, after every layer, k[l] channel characteristics are synthesized and supplied as input to the subsequent layer. The two convolutional operations that make up a dense layer are 3×3 CONV and 1×1 CONV. The dense block of the DenseNet-121 is made up of six of these thick layers. The outcome of every dense block has a depth equal to its rate of expansion.

DenseNet is built on a core convolution and pooling layer. Following a dense block layer, there is an average pooling layer, followed by a classification layer, a change layer, a progress layer, another dense block layer, and lastly, a progress layer. Each dense block contains two convolutions, each with estimated sections of 1x1 and 3×3. The repetition of this occurs frequently in dense block 1, multiple times in dense block 2, 24 times in dense block 3, and lastly 16 times in dense block 4. A standard convolution (1×1) is used to extract the features from each thick layer, and a 3×3 convolution is used to reduce the feature depth/channel count. After every dense block, the total amount of feature maps equals the number of input features.

4.3 Feature selection

After feature extraction, we utilize Enhanced Whale Optimization Algorithm (EWOA) to select the features. The humpback whales' bubble-net feeding technique, served as the model for the whale optimization algorithm. Three strategies were used by the whale optimization algorithm to mathematically represent this behavior: spiral bubble-net assaulting, encircling prey, and looking for prey. Consider a population of whales called $X^t=\left(X_1^t, X_2^t, \ldots, X_N^t\right)$ where each vector $X_i^t=\left(X_{i, 1}^t, X_{i, 2}^t, \ldots, X_{i, D}^t\right)$ specifies the location of the ith whale in iteration t. A matrix $X^1$ is constantly initialized in the issue space in the first iteration (t = 1) and for the subsequent iterations (t > 1), $X^1$ is updated utilizing the three strategies of surrounding prey, searching for prey, and spiral bubble-net attacking. Additionally, the coefficient vector $A_i^t$ is considered when a whale is deciding between two approaches to hunting and circling its target. Using these ideas $X_i^{t+1}$, Eq. (3) is used to determine the $i^{t h}$ humpback whale's position. The probability rate $\rho_i^t$ is a random value between intervals $(0,1)$, the coefficient vector $A_i^t$ is summed utilizing Eq. (4), and $a_i^t$ is computed utilizing Eq. (5) to decrease linearly from 2 to 0 for iterations.

$X_i^{t+1}= \begin{cases}\text { Encircling prey } \left(\rho_i^t<0.5\right) \text { and }\left(\left|A_i^t\right|<1\right) \\ \text { Search for prey } \left(\rho_i^t<0.5\right) \text { and }\left(\left|A_i^t\right| \geq 1\right) \\ \text { Spiral bubble net attacking } \quad \rho_i^t \geq 0.5\end{cases}$         (3)

$A_i^t=2 \times a_i^t \times rand -a_i^t$         (4)

$a_i^t=2-t \times\left(\frac{2}{\text { MaxIt }}\right)$              (5)

4.3.1 E-WOA (Enhanced whale optimization algorithm)

The E-WOA, which adds a pooling mechanism and 3 efficient search tactics called migrating, preferential selection, and enriched surrounding prey, enhances the evaluation of the conventional whale optimization technique.

Definition 1

(Pooling mechanism): Given a matrixPool $=\left(P_1, P_2, \ldots, P_k\right)$ of size with members $P_i=$ $\left(P_{i, 1}, P_{i, 2}, \ldots, P_{i, D}\right)$ that are formed after each iteration using Eq. (6), $X_{\text {brnd }}^t$ is computed utilizing Eq. (9) to construct a random position near the best humpback whale $X_{\text {best }}^t$, and $X_{\text {worst }}^t$ is the worst finding found [23]. In Eq. (6), $\bar{B}_i^t$ is its reverse vector and $B_i^t$ is a binary random vector, where the equivalent values of non-zero components in $B_i^t$ are equal to zero in B _it and the corresponding values of zero components are equal to ones in $\bar{B}_i^t$. To improve diversity, the pooling mechanism uses a crossover operator to combine the unfavorable and promising solutions. A new solution is swapped out for an existing Pool member once the size of the Pool is finished.

$P_i^t=B_i^t \times X_{\text {brnd }}^t+\bar{B}_i^t \times X_{\text {worst }}^t$            (6)

Migrating search strategy: Using Eq. (7), this search approach randomly divides a section of the humpback whale to cover previously unexplored areas and improve exploration. Utilizing Eq. (8), where rand is a shared random amount between 0 and $1, \delta_{\min }$ and $\delta_{\max }$ are the upper and lower bounds of the issue, it is possible to calculate the $X_{r n d}^t$, a random point in the range of the search space. When utilizing Eq. (9) to determine the best humpback whale, $X_{\text {best }}^t$, the $X_{b r n d}^t$ is a random position in the vicinity of $X_{\text {best }}^t$, where $\delta_{\text {best_min }}$  and $\delta_{\text {best_max }}$  are the upper and lower bounds of $X_{\text {best }}^t$.

$X_i^{t+1}=X_{r n d}^t-X_{b r n d}^t$                  (7)

$X_{r n d}^t=\operatorname{rand} \times\left(\delta_{m a x}-\delta_{\min }\right)+\delta_{\min }$            (8)

$\begin{aligned} X_{\text {brnd }}^t=\text { rand } \times & \left(\delta_{\text {best_max }}-\delta_{\text {best_min }}\right)  +\delta_{\text {best min }}\end{aligned}$                 (9)

Preferential selecting search strategy: The canonical WOA's search for prey method has improved exploration potential thanks to the preference selection strategy. The formula for this method is Eq. (10), where $X_i^t$ is the $i^{t h}$ whale's actual position, $P_{r n d 1}^t$, and $P_{r n d 2}^t$ are two random numbers drawn from the Pool matrix in iteration t, $C_i^t$ is specified, and $A_i^t$ is sampled using the Cauchy distribution with the parameter settings. Since the preferential selection search approach is suggested to enhance the whale optimization algorithm exploration capability, sizable step size is required to distribute the whales across many search space regions to find a wide variety of answers. Because of the greater likelihood of creating bigger values, this technique makes advantage of the heavy-tailed Cauchy distribution.

$X_i^{t+1}=X_i^t+A_i^t \times\left(C_i^t \times P_{r n d 1}^t-P_{r n d 2}^t\right)$              (10)

Enriched encircling prey search strategy: Eq. (11), where $P_{r n d 3}^t$  is chosen at random from the matrix Pool and $D^{i t}$ is derived using Eq. (12), is utilized to enhance the encircling prey technique used in the WOA.

$X_i^{t+1}=X_{\text {best }}^t-A_i^t \times D^{\prime t}$                  (11)

$D^{\prime t}=\left|C_i^t \times X_{\text {best }}^t-P_{\text {rnd3 }}^t\right|$                    (12)

The selected features are utilized to classify GI diseases.

4.4 Classification

We propose the new deep-learning SE-ResNet approach for categorizing gastrointestinal disorders. CNN has proven to be the best at computer visual tasks. The SE technique automatically determines each channel's weight through learning and increases the important components. The SE method's architecture is depicted in Figure 2.

2.png

Figure 2. The structure of two operators, excitation and squeeze, make up one block and two SE blocks. Various operations are denoted by variously colored arrows

While CNN-based networks are popular because they outperform alternative architectures in terms of performance, they have drawbacks in that they are difficult to converge and require more memory to train. As a result, ResNets have been proposed as a training method for very deep CNNs. ResNet is a stacked block-wise architecture of the same shape, with each block including direct connections between the result of a lower layer and the inputs of a higher layer. SE-ResNet introduces additional layers called "Squeeze-and-Excitation" blocks within each residual block. These blocks focus on channel-wise feature recalibration, allowing the network to adaptively emphasize important features and suppress less relevant ones. The main components of the Squeeze-and-Excitation block are as follows, the first step is a global average pooling layer, which aggregates information globally from each channel in the feature map. It reduces the spatial dimensions of the feature map to a single value per channel. The pooled features are then passed through two fully connected (dense) layers. These layers learn to model the interdependencies between channels, enabling the network to capture the relative importance of each channel. The output of the excitation step is a set of channel-wise scaling factors that represent the importance of each channel. These scaling factors are used to reweight the original feature map, emphasizing more critical channels and suppressing less important ones.

By using global average pooling, the squeeze operator compresses the given input data to produce a statistical channel. Around the nonlinearity, the excitation processing has a ReLU and 2 completely connected layers. To create a weighted channel, the excitation operator weights the input information [24]. The squeeze-and-excite operators keep the feature channel's size constant. It has frequently been used as the primary classification framework in the task of picture recognition and classification because the residual block of ResNet can effectively leverage shallow categories to acquire more key values. As a result, ResNet was the primary structure utilized in this paper's classification section. Combining the ResNet model and the SE block creates the SE-ResNet module, and Figure 3 depicts the computation method for this module.

3.png

Figure 3. The SE-ResNet module architecture

Excitation and squeeze occur before the summing operation, as illustrated in Figure 3. The SE framework, which can both fully utilize shallow features and additionally reweight every channel of absorb features to improve the classification, is integrated into the residual block of ResNet. The SE-ResNet result is obtained by us as Eq. (13).

$y=F\left(f_{s e}(x),\left(\omega_i\right)\right)+x$               (13)

where, x and y are the SE-input ResNet's outcome, f_se (.) is the method of the SE block, and $\omega_i$ is the network's weight for the $i^{t h}$ input. However, we must specify the scale of the feature picture during the squeeze operation because this has a significant impact on the reweighting value. This research indicates a changeable scale based on the size of the feature channel because the size of every input class picture is not the same. We define the $j^{t h}$ SE-ResNet block's output as in Eq. (14).

$y_j=F\left(f_{s e}\left(x_j\right),\left(\omega_{i j}\right)\right)+x_j$                (14)

$y_j$ is the result of the $j^{t h}$ SE-ResNet architecture, where. We use an optimization approach from the Bald Eagle Search Optimization technique to achieve greater classification accuracy.

4.5 Optimization

For optimization, the bald eagle search optimization algorithm is used. A new meta-heuristic optimization technique called the bald eagle search (BES) was presented in 2020. Bald eagles are at the top of the food chain due to their size. They are sporadic hunters. They can eat any simple, readily accessible protein-rich diet. They prefer to eat fish, particularly salmon, either dead or alive. Because bald eagles have an outstanding vision and the ability to gaze in two directions simultaneously, they can locate fish from a great distance. Their clever social behavior in their hunting method served as the primary source of inspiration for BES. Bald eagles divide their hunting strategy into 3 levels. Choosing space, looking in space, and swooping are these phases. The eagle chooses the area with the most prey throughout the process of choosing the place. The eagle begins looking for prey inside the chosen space during the searching-in-the-space phase [25]. The eagle begins swinging from its ideal position in the swooping phase, which comes last. The best hunting location is then identified. All of the eagle's movements from this point forward are focused on it.

The following is a definition of the bald eagle's hunting system in terms of mathematics:

● Selecting-Space Phase. During this stage, the bald eagle chooses the best location based on the availability of food. Mathematically, this activity is denoted in Eq. (15):

$X_{\text {new }}=X_{\text {best }}+\alpha \times r\left(X_{\text {mean }}-X_i\right)$                (15)

where, $X_{\text {best }}$ denotes the search is chosen based on the best eagle's location, $X_{\text {mean }}$ denotes the average distance between all bald eagle locations (the population mean), $X_i$ denotes the presence of an eagle location, r denotes a random parameter created in the range [0-1], and α denotes a constant parameter.

● Searching-in-Space Phase. In this phase, the bald eagle moves in various directions within the predetermined spiral zone from the previous section in search of prey. Additionally, the ideal location for swooping and prey hunting is chosen. In this stage, the eagle position is updated based on Eq. (16):

$\begin{gathered}X_{\text {new }}=X_i+z(i)  \times\left(X_i-X_{i+1}\right)+p(i) \times\left(X_i\right. \left.-X_{\text {mean }}\right)\end{gathered}$             (16a)

$p(i)=\frac{\operatorname{pr}(i)}{\max (|p r|)}, z(i)=\frac{\operatorname{zr}(i)}{\max (|z r|)}$             (16b)

$\begin{array}{r}\operatorname{pr}(i)=r(i) \times \cos (\theta(i)), z r(i) =r(i) \times \sin (\theta(i))\end{array}$             (16c)

$\theta(i)=\alpha \times \pi \times r 1$             (16d)    

$r(i)=\theta(i)+R \times r 2$             (16e)

where, R is a constant parameter that accepts values between 0.5 and 2, α is a constant parameter that has a value between [0.5, 2], and r1 and r2 are two random parameters.

● Swooping Phase. All bald eagles begin swinging to their discovered prey in this step from the ideal position they achieved in the previous step. At this time, the eagles begin to swing from the best search posture towards their prey, as described in Eq. (17):

$\begin{gathered}X_{\text {new }}=r 3 \times X_{\text {best }}+p l(i) \times\left(X_i-t 1 \times X_{\text {mean }}\right) +z l(i) \times\left(X_i-t 2 \times X_{\text {best }}\right)\end{gathered}$           (17a)

$p l(i)=\frac{p r(i)}{\max (|p r|)}, z l(i)=\frac{z r(i)}{\max (|z r|)}$           (17b)

$\begin{array}{r}\operatorname{pr}(i)=r(i) \times \sin h(\theta(i)), z r(i) =r(i) \times \cosh (\theta(i))\end{array}$           (17c)

$\theta(i)=\alpha \times \pi \times r 3, r(i)=\theta(i)$           (17d)

where the two constant variables are t1 and t2.

The earlier stages are thought to be essential for quickly arriving at a good solution.

Managerial implications for gastrointestinal (GI) diseases are essential for healthcare organizations, medical professionals, and policymakers to address the challenges associated with the prevention, diagnosis, treatment, and management of these conditions. Adequate healthcare infrastructure is crucial for efficiently managing GI diseases. Managers should ensure that hospitals, clinics, and healthcare facilities have the necessary equipment, specialized units, and qualified staff to handle GI-related cases effectively. Managerial efforts should be directed towards raising public awareness about GI diseases, their risk factors, and early symptoms. Educational campaigns can help promote healthy lifestyle choices and encourage individuals to seek timely medical attention. Developing screening programs for high-risk populations can aid in the early detection of GI diseases, such as colorectal cancer. Timely detection can improve treatment outcomes and reduce healthcare costs. Managers can support research and development initiatives to advance the understanding and treatment of GI diseases. Encouraging clinical trials and collaboration with academic institutions can lead to innovative therapies and improved patient outcomes.