An Enhanced Approach to Liver Disease Classification: Implementing Convolutional Neural Network with Attention Layer Gated Recurrent Unit

Colorectal cancer, along with other conditions in the liver region, are typically caused by malignant polyps that form in the colon. Based on statistical data, colon cancer is the most prevalent type in the United States. Fortunately, the proliferation of early detection programs for these tumors and their precursor polyps is helping to mitigate the prevalence of colorectal cancer. It's broadly recognized that adenomatous polyps - benign lesions with dysplastic epithelium and varied malignancy potential - account for over 95% of colorectal malignancies. The progression from adenoma to carcinoma is well understood. Generally, this is a slow process, taking many years to manifest following an incremental accumulation of genetic alterations.

Malignant polyps penetrate the muscularis mucosa but remain confined to the submucosa (pT1). Adenomatous polyps may exhibit high-grade dysplasia and other non-invasive histological features. Such polyps constitute up to 12% of polypectomy series, a figure that is on the rise due to increasingly effective screening programs using colonoscopy. These programs not only prevent colorectal cancer but also treat some advanced polyps. Between 80% and 90% of adenomas are smaller than 1cm, enabling standard snare polypectomy to be executed more easily, especially for pedunculated polyps. Larger lesions are treated with endoscopic mucosal resection (EMR) and endoscopic submucosal dissection (ESD) at specialized centers, allowing for total rather than piecemeal excision. Full removal enables a more comprehensive histological examination, thus serving as the initial step in malignant polyp management. However, in clinical practice, scenarios often differ. A patient typically presents for examination after a resected polyp, initially thought to be benign during endoscopy, is subsequently found to have invasive adenocarcinoma upon pathological evaluation. This situation may be further complicated if the polypectomy site wasn't marked during the initial endoscopy, hindering endoscopic re-evaluation and making colon segment identification unreliable if definitive resection is required. The consulting physician will need to stratify risk by considering potential residual or recurrent disease, lymph node metastases, and operative risk. Current data indicates that this remains a contentious topic requiring a multidisciplinary approach. The prognostic features of malignant polyps will most significantly alter this risk profile. Effective management strategies will also be explored [1].

Colon cancer is the second leading cause of death worldwide and the third most common cancer. Expert physicians are necessary for early diagnosis of this malignancy [2]. Gastrointestinal diseases, such as Crohn's disease, cancer, polyps, bleeding, and ulcers, are prevalent. Researchers in the Computer Vision domain employ various computer-based techniques for early-stage liver cancer diagnosis, but the handcrafted features used in the process sometimes lead to misclassification. Feature extraction and lesion segmentation receive significant attention in image processing and Computer Vision across various applications, including agriculture, medical imaging, and surveillance [3]. Early stages of small liver lesions often go undetected and can lead to fatal conditions. Hence, it's crucial to implement computer-based methods to aid doctors in effectively diagnosing and treating patients [4, 5]. Quantitative features, consisting of both visible non-texture features and invisible texture features, are extracted from input data and applied to the classification model. Statistical methods or machine learning techniques are utilized to build a robust classifier for the classification of gastrointestinal diseases [6].

Existing research has leveraged various feature extraction methods such as Log filter bank, texture features, Scale Invariant Feature Transform (SIFT), discriminative joint features, and color features [7, 8]. Feature reduction techniques employed include Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), while classifiers such as K-Nearest Neighbor (KNN), Naïve Bayes, Support Vector Machine, and Artificial Neural Networks (ANN) have been applied for robust classification [9]. Backpropagation algorithms have been utilized to extract clinical features, and deep learning techniques have enabled models to handle a large number of training images. This capability allows machines to diagnose newly acquired clinical images based on a database of clinical features. Convolutional Neural Networks (CNN) are deep learning systems that accurately emulate the structure and function of brain neurons on computers [10, 11].

The CNN-AGRU model is proposed for liver disease classification to enhance efficiency. The input images undergo preprocessing for image augmentation and enhancement. The AGRU model performs the classification process, while the CNN model extracts features from the input images. The performance of the CNN-AGRU model in liver disease classification is evaluated and compared with existing machine learning and deep learning techniques.

Machine learning models, due to their lack of feature learning and difficulty with data imbalance, perform less effectively when classifying liver diseases. The proposed model was compared with the Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors Algorithm (KNN) methods using machine learning techniques. Due to overfitting issues, SVM, KNN, and Decision Tree present unstable performance compared to Random Forest. Performance indicators compared include accuracy, sensitivity, and specificity. The accuracy rate for CNN-AGRU is 98.2%, while the rates for the DT, RF, SVM, and KNN models are 74.3%, 82.1%, 85.1%, and 86.2%, respectively. Similarly, CNN-AGRU's sensitivity value is 98.5%, while the DT, RF, SVM, and KNN models' respective sensitivity values are 76.2%, 83.2%, 86.7%, and 85.3%. Moreover, CNN-AGRU has a specificity value of 99.2%, whereas the DT, RF, SVM, and KNN models achieved specificities of 77.1%, 85.7%, 87.9%, and 86.8%, respectively.

The overfitting problem of the CNN model in the deep learning framework lowers classification performance. RNN models perform poorly in classification due to the loss of critical knowledge acquired early in the training process. The LSTM model's network updates also lose crucial data for classification due to the vanishing gradient problem. Meanwhile, the GRU approach has an overfitting issue in classification and discards new irrelevant features to improve learning. The CNN-AGRU model provides energy vector normalization to extract distinct features from the input data. The attention layer of the GRU better identifies pertinent features by learning the context vector of the spatio-temporal model. The accuracy of the CNN-AGRU model is 98.2%, whereas that of the RNN, LSTM, GRU, and CNN is 88.2%, 91.8%, 92.3%, and 95.1%, respectively.

Unlike conventional methods, the CNN-AGRU model discards recently learned irrelevant information, while the standard GRU model selectively identifies relevant features. The use of collaborative feature learning enhances the performance of the CNN-AGRU model when learning features with GoogleNet, ResNet-50, AlexNet, and TAS-RF. The CNN-AGRU model learns a context vector, which is responsible for selecting pertinent features for spatiotemporal input classification. A distinguishing feature of the CNN-AGRU model is its usage of energy vector normalization, which aids in learning distinctive features and addressing imbalances in data. CNN-based models using AlexNet, ResNet-50, and GoogleNet are limited by an overfitting issue. The LSTM-CNN model's performance is hindered due to increased overfitting, and the enhanced RCNN model is less efficient at feature learning. The CNN-AGRU model has an accuracy of 98.2%, while the accuracy of the GoogleNet, ResNet-50, AlexNet, and TAS-RF models is 96.7%, 95%, 97%, and 86.2%, respectively.

Convolutional Neural Network-Attention Layer Gated Recurrent Unit (CNN-AGRU) for Liver Disease Classification offers several advantages, including:

High Accuracy: The model has demonstrated excellent accuracy in classifying liver disease cases, assisting healthcare professionals in formulating more precise diagnoses and treatment plans.

Efficient Feature Extraction: The use of a Convolutional Neural Network (CNN) enhances the extraction of significant features from liver images, thereby improving disease classification accuracy.

Robustness: The method exhibits resilience against variations in liver image brightness and contrast, leading to more consistent and reliable diagnoses.

Attention Mechanism: The attention mechanism helps the model focus on the most relevant parts of the image, boosting the classification accuracy.

Attention Layer Gated Recurrent Unit (GRU): The use of a GRU helps the model capture temporal dependencies in image data, aiding in pattern recognition and accurate predictions [12].

Automated Diagnosis: The model can automatically diagnose liver diseases, reducing the workload of medical professionals and improving patient outcomes by facilitating early detection of disorders.

Non-invasive: As this method relies on routine medical imaging for diagnosis, it is non-invasive, avoiding the need for additional invasive tests or procedures.

The structure of this paper is as follows: Section 2 presents recent deep learning techniques for classifying liver diseases, while Section 3 describes the CNN-AGRU model. Section 4 provides details on the implementation of the CNN-AGRU model, and Section 5 presents the results of liver disease classification using the model. Finally, the conclusion of this research study is presented in Section 6.

The performance of the suggested model was assessed using the liver tumour dataset. The pre-processing for categorization used image augmentation and enhancement techniques. The feature extraction was done using Convolutional Neural Network (CNN) from the pre-processed image. The extracted features were applied to the AGRU method for the classification of Liver disease. Figure 1 shows the proposed CNN-AGRU model in Liver disease classification.

3.1 Pre-processing

Images of the Liver were pre-processed before applying to CNN models to resize and enhance the images. Color constancy was scaled in the image, and the image size was changed to 244×244 pixels for ResNet-50 and GoogleNet and 227×227 pixels for the AlexNet model. The three RGB channels mean it was calculated for Liver images, and an average filter was used to enhance the images. The average filter measures each pixel's average value with its neighbors and replaces it, and this process is carried out for the whole image.

CNN model is highly based on data volume, and training data in a more extensive set generates a model for promising results. The data augmentation method improves the classification accuracy for the CNN model, and data augmentation was applied due to the lack of medical images. The data augmentation also balances the dataset for some images between classes. The rotation, shifting, zooming, and flipping operations were used in augmented training data in this research.

1.png

Figure 1. The CNN-AGRU model in liver diseases classification

3.2 Convolutional Neural Network

Convolutional Neural Network (CNN) is based on Deep Learning (DL) and Neural Networks (NN) [17, 18], which is suitable for handling 2D images. CNN model consists of Fully Pooling Layers (PL), Convolutional Layer (CL), and Connected Layers (FCL). Because CNNs acquire features through training and dramatically shorten the time for feature design or choosing the most distinctive features, they perform better than other machine learning techniques like Decision Tree, SVM, and Naive Bayes.

Convolution is an essential process in CNN, and the convolution layer is an important layer in CNN that performs a 2D convolution process on input and passes in kernels. Each Convolution Layer kernel’s weights are randomly initialized, and the network training loss function is updated at each iteration. Final Kernels learn some types of patterns in input images.

CNN model consists of non-linear activation function, stack, and Convolution. Consider X as an input matrix and Convolution Layer has an output O and a set of kernels F_jexists $\forall_j \in[1, \cdots, J]$, then convolution output is defined in Eq. (1).

$C(j)=X \otimes F_j, \forall_j \in[1, \ldots, J]$           (1)

The dot product of inputs and filter is defined using convolution operation $\otimes$.

In Eq. (2) a new 3D activation map is applied using C(j) the activation map.

$D=S(C(1), \ldots, C(J))$           (2)

where, $J$ stands for the total number of filters and $\mathcal{S}$ stands for the pile operation's channel direction.

Eq. (3) provides the 3D activation map D final output activation map and Non-Linear Activation Function (NLAF).

$O=N L A F(D)$           (3)

The output, filters, and input are three important matrixes with sizes S, as given in Eq. (4).

$S(x)=\left\{\begin{array}{cc}V_1 \times Q_1 \times H_1 & x=X \\ V_K \times Q_K \times H_K & x=F_j, \forall_j \in[1, \ldots, J] \\ V_0 \times Q_0 \times H_0 & x=0\end{array}\right.$             (4)

where, activation map of channels, height size, and width are denoted using three variables $(V, H, Q)$, respectively. The output, filter and input are denoted using subscripts $I, K$, and $O$. It has two equalities: $H_K$ indicate filter channel which is equal to the input channel $H_I\left(H_I=H_K\right)$, the second number of filters $J$ is equal to the output channel $H_O$ that is $H_O=J$.

Eq. (5) and Eq. (6) measure values of $\left(V_O, Q_O, H_O\right), A$ the stride, and padding denotes $B$.

$V_0=1+f_{f l}\left[\left(2 \times B+V_1-V_K\right) / A\right]$          (5)

$Q_0=1+f_{f l}\left[\left(2 \times B+Q_1-Q_K\right) / A\right]$                  (6)

where, the term "floor function" is used as $f_{f l}$.

Eq. (7) selects rectified linear unit (ReLU) and NLAF is denoted as $\sigma$.

$\sigma_{R e L U}\left(d_{i j}\right)=\operatorname{ReLU}\left(d_{i j}\right)=\max \left(0, d_{i j}\right)$              (7)

where, activation map $D$ elements denote $d_{i j} \in D$. Sigmoid (SM) function and traditional Hyperbolic Tangent (HT) have considerable performance, and ReLU is a popular Non-Linear Activation Function (NLAF), as in Eq. (8) and Eq. (9).

$\begin{aligned} \sigma_{H T}\left(d_{i j}\right)=\tanh & \left(d_{i j}\right) =\left(e^{d_{i j}}-e^{-d_{i j}}\right) /\left(e^{d_{i j}}+e^{-d_{i j}}\right)\end{aligned}$        (8)

$\sigma_{S M}\left(d_{i j}\right)=\left(1+e^{-d_{i j}}\right)^{-1}$        (9)

ReLU is one-sided that is more plausible biologically than $\sigma_{H T}$. The CNN model is shown in Figure 2.

2.png

Figure 2. The CNN model

3.3 Attention Gated Recurrent Unit

Generally, the RNN suffer from vanishing and exploding gradient problem in the classification. RNN has a limitation of exploding gradient, and the term ‘long-term’ of gradient grows exponentially faster than short term. GRU [19] and LSTM [20] are RNN types, and unlike CNN, the RNN model backward connection affects the model accuracy, and the LSTM model is developed to tackle this problem. The LSTM model is designed for temporal features handled in long-range dependency, and cell blocks are present in LSTM internal structure. The hidden and cell states are transferred from one block to another, and that memory block is used to remember through gates.

A complex non-linear function is handled in sequence-to-sequence architecture, and encoder-decoder GRU networks with attention models jointly promote features. Specific spatial-temporal inputs are linked with a context vector at each time step t. Most relevant input features are adopted in the attention mechanism, as in Eq. (10).

$e_{j t}=\tanh \left(W_a s_{t-1}+U_a h_{j t}\right)$            (10)

where, weight matrices are denoted as $U_a$ and $W_a$. The normalized weight of the energy vector is given in Eq. (11):

$\alpha_{j t}=\left(\frac{\exp \left(e_{j t}\right)}{\sum_{l=1}^T \exp \left(e_{l t}\right)}\right)$       (11)

where, input sequence of local region and target symbol denotes the alignment of attention probability in $\sum_{j=1}^T \alpha_{j t}=1$, and $a_{j t}$. Attention probabilities of the encoded hidden state are measured based on weighted sum at each step at time step t of the context vector.

Output prediction of different encoded input variables used to represent the vector. GRU model of sequence to sequence with attention to measure input sequence elements at required attention.

LSTM has three gates such as input, output, and forget. GRU model has two gates an update and a reset gate. The update gate checks earlier cell memory to keep it active, and the Next cell combines with the preceding cell memory, which is carried out by the reset gate. LSTM input and forget gate are merged to the update gate, and the reset gate of the hidden state is directly applied. The GRU cell of general equations is given in Eq. (12)-Eq. (15). GRU of multi-layer considers faster training based on a smaller number of parameters.

$z_t=\Theta\left(W_z \cdot\left[h_{t-1}, x_t\right]+b_z\right)$           (12)

$r_t=\Theta\left(W_r \cdot\left[h_{t-1}, x_t\right]+b_r\right)$                   (13)

$\widehat{h}_t=\tanh \left(W_h \cdot\left[r_t \cdot h_{t-1}, x_t\right]+b_h\right)$           (14)

$h_t=\left(\left(1-z_t\right) \cdot h_{t-1}+z_t \cdot \hat{h}_t\right)$         (15)

The sequence representation of CNN is measured using multi-layered GRU for its effective sequence learning. Spatial feature extraction of CNN layers is measured from refined input data, and multi-layer GRU is fed into the model. Two CNN layers of the ReLU activation function are applied and followed by a kernel filter with the size of 2 in 1×16 and 1×8, which were applied for the first and second layers, respectively. GRU layers are applied with spatial extracted features. Two GRU layers used temporal characteristics, and a dense layer carried out the prediction. The GRU model is shown in Figure 3.

3.png

Figure 3. GRU model

3.4 Managerial implications

The proposed method can assist doctors and medical workers in more precisely and rapidly diagnosing liver illnesses, leading to better patient treatment outcomes. Deep learning techniques can also assist in uncovering trends in medical data that are not visible to the naked eye, resulting in more successful treatment strategies.

Using automated classification systems can reduce medical personnel workloads, allowing them to focus on other areas of patient care. This can result in improved patient outcomes and increased job satisfaction among healthcare professionals.

Automated classification systems can also minimize manual diagnosis and treatment expenses, which are significant in healthcare organizations. Deep learning techniques can assist in identifying potential health risks early on, avoiding the need for more expensive and intrusive procedures later on.

As with any program that handles patient data, it is critical to safeguard the system's security and patient privacy. Healthcare organizations must establish adequate security measures to ensure that patient data is not compromised or misused.

The proposed strategy is a foundation for future automated medical diagnosis and therapy studies. Building on this work, healthcare organizations and researchers can create more advanced deep-learning models for different medical applications, leading to better patient outcomes and healthcare delivery.

The CNN-AGRU model is proposed for Liver disease classification to increase efficiency. The pre-processing of image enhancement and augmentation were carried out in input images. The CNN model extracts the features from the input images, and the AGRU model performs classification.

The CNN-AGRU model is tested on Liver disease classification and compared with machine learning models, as given in Figure 4 and Table 1. Machine learning models perform less in Liver disease classification due to a lack of feature learning and an imbalanced data problem. The Decision Tree and Random Forest provide unstable performance due to the overfitting problem. The KNN model has outlier restrictions, but the SVM model struggles with data imbalance. The CNN-AGRU model offers a standard feature learning method that improves the feature learning efficiency in the GRU model. The context vector learning in the attention layer of GRU helps to learn the context vector of a spatio-temporal model.

The CNN-AGRU model is evaluated in Liver disease classification and compared with the deep learning model, as given in Figure 5 and Table 2. The CNN model's overfitting limitation degrades the performance of classification RNN models do less well in classification because they lose the essential information learned early in the training process. The vanishing gradient problem in the LSTM model causes network updates to lose important information for categorization. GRU method washouts the new irrelevant features to improve the learning and has an overfitting problem in classification. The CNN-AGRU model provides energy vector normalization to learn the unique features from the input data. The attention layer in GRU learns a context vector related to the spatio-temporal model to select the relevant features and improves efficiency. The CNN-AGRU model has 98.2% accuracy, and GRU has 92.3% accuracy.

4.png

Figure 4. CNN-AGRU and standard classifiers in liver disease classification

5.png

Figure 5. CNN-AGRU and deep learning models in liver disease classification

Table 1. CNN-AGRU model and standard classifier comparison

Table 2. CNN-AGRU model and deep learning model comparison

5.1 Comparative analysis

The existing methods in related works were compared with the CNN-AGRU model to show its efficiency.

The CNN-AGRU model is compared with existing Liver disease classification methods, as shown in Figure 6 and Table 3. Unlike the average GRU model, which chooses the pertinent features, the CNN-AGRU model discards newly learned irrelevant information. The CNN-AGRU model learns features collaboratively, which enhances the model's feature learning performance. When classifying spatiotemporal input, the CNN-AGRU model learns a context vector that chooses the pertinent features. The CNN-AGRU model has energy vector normalization that helps to learn unique features and solves imbalanced data problems. The CNN-based models of GoogleNet, ResNet-50, and AlexNet have the limitation of overfitting issues. The LSTM-CNN model increases the overfitting in the model and degrades the performance. The improved RCNN model has lower efficiency in feature learning. The CNN-AGRU model has an accuracy of 98.2%, and the Google-Net model has 96.7% accuracy.

Table 3. CNN-AGRU model and existing method comparison

6.png

Figure 6. CNN-AGRU and existing methods in Liver disease classification