An Accurate Pneumonia Detection Scheme Using Hybrid Algorithms

The accuracy of doctors' diagnoses and treatments of pneumonia are directly impacted by the quality of medical images created from such cases, making it a critical component of modern medicine. Achieving accurate diagnoses of pneumonia cases is difficult as a result of the issues of low contrast and resolution of the created medical images, which directly affects the accuracy and speed of doctors' examinations. De-noising of medical images while preserving fine details is a challenging issue in image processing. Therefore, there is a need to enhance medical image quality to accurately and represent the information related to an illness [1, 2]. The purpose of medical pneumonia image enhancement methods is to improve the visual quality of pneumonia images. These methods aim to improve image features without changing the underlying information they contain. There are various methods employed to achieve this, including blur removal and de-noising techniques [3-5].

Pneumonia image enhancement in the literature can be categorized into two domains- spatial and frequency domains. Pneumonia image enhancement in the spatial domain involves manipulating the image pixel values, whereas, in the frequency domain, enhancement is achieved through a transform approach applied to the pneumonia images. Several techniques for enhancing pneumonia images rely on spatial processes that operate on individual image pixels. These algorithms aim to generate pneumonia images that are more suitable and improved compared to the original input image [6].

The fractional differentials theory (in mathematics) deals with differentials of arbitrary order; it should be noted that fractional differentials (FD) have similarities to integral differentials [6, 7]. When compared to traditional integral differential techniques, the employment of FD in image processing improves edges, preserve smooth areas, and improve the visibility of texture features [6, 8]. Improved contrast and clarity are seen in medical images of pneumonia that have been processed using fractional differentials. In the conventional FDs, textures, edges, and smooth image areas are processed using a consistent fractional order, but this approach has limitations. While edges can be efficiently enhanced by higher fractional orders, smooth areas and poor textures may be overlooked. Lower fractional orders, on the other hand, may weaken the edges but may keep weaker textures and smoother areas. As a result, improving images in practice is difficult. Researchers have created improved and conventional fractional differential algorithms for digital image processing in order to address these issues [9, 10]. Adaptive fractional derivatives have also been investigated in earlier research to address image de-noising issues [11, 12]. For instance, six FD masks and the YiFeiPu-2 operator were developed by Pu et al. [13] which achieved better convergence and precision owing to the fixed human assigned values of the fractional orders in the earlier mentioned differential operators.

Numerous research in the area of computer-aided design-based artificial intelligence (CAD-AI) has focused on the diagnosis of pneumonia infections, including classification, detection, and segmentation of suspected pneumonia. However, these studies have faced limitations due to issues related to the pneumonia image dataset. Additionally, high accuracy in pneumonia disease classification has not yet reached the optimum level [5, 14]. Because the symptoms of the disease are not readily apparent on CT or X-ray scans, the objective and automated identification of pneumonia poses a significant problem in medical imaging. Furthermore, addressing this task is of utmost importance as millions of individuals succumb to pneumonia annually [15].

Through a brief survey of previous studies, deep neural networks have demonstrated remarkable promise for image classification. Nonetheless, real-valued data provides the foundation for the majority of recent research on image classification frameworks. The real-valued CNN, according to study presented by Yoon and Kang [16], is unable to accurately capture the relationship between multi-image channels. the need to devise a new approach that addresses the shortcomings of the previous proposed methods becomes clear. Therefore, this paper presents a novel and efficient RN50 CNN-AFDA method that has been applied to automatically detect pneumonia from chest X-ray images, as well as analyze and learn the patterns that indicate pneumonia from chest X-rays. The proposed scheme has four main phases; first, a combined approach of ResNet50 and Transfer Learning algorithms are used to enhance the collected image's quality and make it appropriate for the proposed model (i.e., data augmentation and normalization). Second, the adaptive fractional differential function is utilized to determine the optimal fractional order for each pixel. This function is constructed by taking into account the characteristics of various image areas. Third, the fractional differential mask is modified to include an ideal fractional order to process the matching image pixel. Finally, a deep learning approach based on the Adam optimizer algorithm is utilized to lower the learning rate, fasten the convergence, and enhance the model accuracy. The rest of this article is arranged thus: An overview of the related theories to fractional difference and the implementation of fractional differential masks is presented in Section 2. Section 3 presents the adopted methodology in the present study. In Section 4, an extensive analysis of the conducted experiments and subsequent comparisons was presented. Finally, the study is concluded in Section 5.

This section explains the methodology based on RN50 CNN-AFDA, which is used to identify pneumonia infection in X-ray images. The methodology involves several steps; first is the acquisition of pneumonia images; second, several actions were performed during the data preprocessing phase. Third, the proposed RN50 CNN-AFDA method was designed and created. Fourth, the proposed scheme was evaluated and tested to compare its achievements with existing studies. Figure 1 depicts the flowchart of the suggested work.

1.png

Figure 1. The flowchart of the suggested scheme

The primary measures taken to execute the proposed scheme are depicted in Figure 1. In the proposed model, the output from each step is used in the subsequent step as input. Any DL model must first collect a sufficient dataset based on the goal that has to be accomplished. Therefore, the most appropriate dataset for this study is the 5228 medical X-ray dataset, which consists of chest X-ray images that are categorized as either pneumonia infection or normal. Once the dataset has been acquired, a crucial step is data pre-processing, which involves preparing the data for use by the specific model. The data is utilized for both training and testing the ResNet50 model in this case. Once the model is developed, it undergoes a training process before it can be tested on new data. The models’ performance is then evaluated based on this testing. Model evaluation often involves the use of multiple metrics to assess its overall performance efficiency.

3.1 Dataset description

As previously mentioned, it is important to utilize an appropriate dataset when attempting pneumonia detection from chest X-rays. The pneumonia X-ray dataset is a modest in size, it has 3 medical folders (COVID-19 images, Pneumonia images and Normal images) these three folders containing Chest X-ray (CXR) Images. All images are preprocessed and resized to 256×256 PNG format. The used dataset has 5228 medical images, 1626 images for COVID-19, and 1800 images for Pneumonia and finally 1802 as normal images. These images were obtained from the Mendeley Data repository [20]. The dataset contains two categories of images: pneumonia and normal. The utilized dataset helps the medical community and researcher to detect and classify COVID19 and Pneumonia from Chest X-Ray Images using Deep Learning techniques.

2.png

Figure 2. X-ray training images (pneumonia and normal categories)

3.png

Figure 3. Random images from the X-ray dataset and their corresponding labels as (pneumonia or normal) [7]

Each image in this dataset is labeled as one of two categories. This means that both training images and testing images, as well as validation images, contain examples from both categories. The distribution of the training images across the two groups is detailed in Figure 2.

Furthermore, each category includes X-ray images for the training, testing, and validation processes. In the pneumonia category, there are 1590 images for training process and 202 for the testing phase; there are also 8 images for validation. The Normal category has 1568, 226, and 8 images for training, testing, and validation purposes, respectively. Figure 3 displays an example of the dataset images.

The chosen dataset is quite suitable for the pneumonia detection task; it is relatively big and diverse. Furthermore, there are actual images labeled and verified to clearly depict X-rays of patients with diseases, as well as normal X-rays. Also, the dataset includes high-resolution images that help the DL model to create improved results for the given project. Thus, the availability of this high-quality dataset will enable a very accurate method of diagnosing pneumonia in chest X-rays using the proposed RN50 CNN-AFDA method.

3.2 Data pre-processing

This is an essential step in any ML or DL-based project. Preparing the data is also an important stage which includes cleaning, transforming and normalization of the data that in this case would be most suitable to feed to the ResNet50 CNN model selected. This process improves the quality of the data which makes it fit for the model to use. It also enhances the quality of results obtained, eliminates descriptive errors, and prejudice, and reveals useful information about the data. The first among the preparatory stages of data pre-processing is to read in the data. Images in the dataset are then read from disk employing graphical libraries such as OpenCV or PIL in Python. Furthermore, all the images are resized to the dimension of 224×224 to satisfy the requirements of the further proposed RN50 CNN-AFDA approach. The algebraic relation expressed below denotes the process of resizing:

$f: I->I^{\prime}$               (6)

where, I = initial image, I' = resized image.

The next procedure of pre-processing is data augmentation, and it is a method used in enhancing and creating new data from the pre-existing dataset. This technique helps to add more variations to the image and contribute to the acquisition of the given model’s robustness. Augmentation concerns operations such as rotation, crop and mirror images to the existing set of images. Here is the explanation of the rotation process:

$x^{\prime}=x \cos \theta-y \sin \theta$             (7)

$y^{\prime}=x \sin \theta+y\cos\theta$                (8)

where, θ = angle of rotation, x, y = original coordinates of a pixel, x’, y’ = post-rotation coordinates of a pixel.

Conversely, flipping can be applied on the images to flip them horizontally, vertically or side wise. The function that follows represents the horizontal flipping:

$f:(x, y)->(W-x, y)$                (9)

where, W = width of the image.

Lastly, data normalization is done on the dataset to help quicken the models’ convergence during training; it also reduces the possibility of local optima entrapment. Data normalization is often performed to rescale data to a common scale, usually between 0 and 1 or -1 and 1, without distorting the differences in the ranges of the data. This is typically done by subtracting the mean μ and dividing by the standard deviation σ of the pixel values in the dataset:

$p^{\prime}=(p-\mu) / \sigma$               (10)

Such that p is the original pixel value, and p’ is the normalized pixel value.

Overall, all these pre-processing steps contribute to preparing the dataset for RN50 CNN-AFDA training, and to the effectiveness of the training process.

3.3 The proposed scheme

Three main methods were used in the proposed scheme; these are the ResNet50 CNN method, the transfer learning method, and the AFDA algorithm. The combination of ResNet50's deep architecture and transfer learning makes it a robust solution for pneumonia detection, improving accuracy while reducing the need for large datasets or extensive computational resources. There are several advantages for combining ResNet50's deep architecture and transfer learning, these are:

Efficiency: Training deep models from scratch on medical datasets, which are often small, is time-consuming and prone to overfitting. Transfer learning helps in leveraging the general patterns learned from large-scale datasets, speeding up the process.

Generalization: ResNet50, with transfer learning, can generalize better to pneumonia detection by adapting pre-learned knowledge to the specific characteristics of medical images.

Resource saving: A pre-trained model requires significantly fewer computational resources compared to training a deep model from scratch, making it ideal for tasks like pneumonia detection in resource-constrained settings.

3.3.1 ResNet50 CNN

ResNet50 was developed by He et al. [21] for addressing the issue of gradient disappearance and explosion encountered during model training. ResNet50 CNN offers a more straightforward training process that eliminates the possibility of network degeneration. The significance of the ResNet50 algorithm with 152 layers is demonstrated on the on the ImageNet dataset where it achieved only 3.57% error. In the traditional ResNet50 model, the convolutional layers of VGG's model serve as the basis, where there are two similar 1×1 convolution layers in each residual block. After these convolution layers comes the batch normalization layer, followed by the rectified linear unit (ReLU) activation function. The Softmax function is then used to guarantee normalization. An overview of the typical ResNet50 architecture is presented in Table 1.

Table 1. The ResNet50 network parameters

Layer Name	Output Size	50-Layer
Conv1	112×112	7×7, 64, stride 2
Conv2.x	56×56	3×3 max pool, stride 2
		1×1643×3641×1256×3
Conv3.x	28×28	1×11283×31281×1512×4
Conv4.x	14×14	1×12563×32561×11024×6
Conv5.x	7×7	1×15123×35121×12048×3
	1×1	Average pool, 1000-d fc, Softmax
FLOPs		3.8×109

3.3.2 Transfer learning

The practice of using knowledge obtained from one or more sources to help extract features from a target domain is known as transfer learning (TL). This is especially helpful if there aren't many data samples available. This study requires “a TL-ResNet50 model that has already being trained on the ImageNet dataset. The TensorFlow framework was utilized to load the model. While the other layers of the model remained frozen, the output feature parameters were adjusted. The TL-ResNet50 model was trained on the ILSVRC 2012 dataset that contains more than a million 224×224-pixel images of natural scenes. ReLU is a nonlinear activation function that is widely used in the deep learning sector. The ReLU activation function is highly popular due to its ability to provide a straightforward non-linear transformation, effectively addressing the issue of gradient disappearance. The ReLU function is defined [22].

$\operatorname{ReLU}(X)=\max (x, 0)$               (11)

3.3.3 Adaptive estimation

The conventional method of updating weights as used in machine learning often referred to as the random gradients assigns the same Alpha throughout the training session to all weights. However, this technique may not be very efficient especially due to the fact that the learning rate normally needs to be adjusted at various instances to ensure optimal convergence. Whereas Adam algorithm [4] produces the adaptable learning rates for required parameters in the procedure, the Adam algorithm is well known to be easy to implement, to be numerical efficient, and to use little memories. Thus, the performance of the Adam method is higher than the RMSProp algorithm, and it is defined as:

$v_t \leftarrow\left(1-\beta_2\right) \sum_{i=1}^t \beta_2^{t-i} \cdot g_i^2$                 (12)

where, $v_t$ is updated as a weighted sum of past squared gradients. $\beta_2$ is a hyperparameter that controls the decay rate of the past squared gradients (usually close to 1, like 0.999).

To enhance the model's accuracy and convergence, this work combined the Adam optimizer with a decreasing learning rate. A better learning rate was achieved during the initial training phase by halving the models’ learning rate after 10 training cycles; this process also improved the network convergence. The later stage of the training used a lower learning rate to enhance convergence towards optimal solutions.

3.3.4 Cross-entropy loss function (CELF)

The CELF is a widely used approach [23] for improving the models’ prediction accuracy; it examines the variations between the predicted and actual probability distributions. The CELF approach mostly focuses on the probability of a result occurring; and as long as the result is significant, the outcome of the prediction is taken as precise. The accuracy and training speed of the model is improved significantly because the CELF technique can rapidly alter the weight (W) and deviation (B). The CELF is defined as seen below:

$L\left(y^{(i)}, y h a t^{(i)}\right)=-\sum_{j=1}^q y_j^{(i)} \log y h a t_j^{(i)}$                (13)

where, $y_j^{(i)}$ vector = element of the $y^{(i)}$ vector that is not 0 or 1; $y^{(i)}$ = discrete value of the sample i category.

3.3.5 TL-ResNet50 CNN model

The TL-ResNet50 CNN model consists of 5 convolutional kernels and 50 neural network layers. Fundamental convolutional operations like convolution (CONV), batch normalization (BN), and rectified linear unit (ReLU) are used in each layer of the model. The training set is simply handled in a single computation for the stage 0 convolutional layer. The bottleneck structure used in Stages 1 through 4 is made up of numerous similar residual components. In stage 1, residual blocks are included to improve the effectiveness of feature extraction while addressing gradient disappearance and explosion issues. The model starts to increasingly extract more sophisticated and abstract features at stages 2, 3, and 4. Target categorization tasks are completed at stage 4 of the process.

4.png

Figure 4 illustrates the design of the proposed TL-ResNet50 model. A 1×1 convolutional layer is part of the ResNet50 core component, which is shown as the "Bottleneck basic residual module (BTNK1 and BTNK2)" on the right side of the image. This basic module was included with the intention of supporting deeper networks. Assuming there are 256 channels in the input, the number of channels will be decreased to 64 by the 1×1 convolutional layer. A 3×3 convolutional layer is then used to boost the number of channels to 256. Reducing the amount of network parameters by using a 1×1 convolutional layer eases the computational burden on deeper networks.

In general, CNNs consist of various components and one of the most popular CNN networks is ResNet. In ResNet, the basic block is made up of two parts: a convolutional layer and a shortcut connection, also known as an identity connection. The accuracy of the model increases directly with the number of modules in the network. Throughout the years, various modifications and improvements have been made to the ResNet networks. This section introduces the basic network, the original ResNet, and ResNet with BN. Matrix multiplication can be used in the case of ResNet since its convolution operation is linear. The matrix multiplication can be described as follows: If the input vector is $x \in \mathbb{R}^{n^2}$, the weight matrix is $W \in \mathbb{R}^{n^2 \times n^2}$, and the characteristic output obtained after the convolution operation is $y \in \mathbb{R}^{n^2}$; the convolution process can be written as $y=W^T x$.

5.png

Figure 5. A general diagram of the basic blocks within ResNet, and with BN [24]

(1) Basic block

Figure 5(a) presents the basic block, which serves as the backbone of the convolution network. In this block, the input data undergoes convolution to extract features. To differentiate between them, the same basic block is introduced to the ResNet basic network (b) as well. The only distinction between the basic block (a) and the ResNet block (b) is the presence of a shortcut connection in (b). In the figure, $x_l$ is the input feature of the l-th module. $f: x_l \rightarrow x_{l+1}$ is the mapping of the input vector $x_l$ to the output vector $x_{l+1}$ in the basic network. Consequently, the output of this layer is:

$x_{l+1}=f\left(x_l,\left\{W_l^{(i)}\right\}\right)=W_l^{(2)} \sigma\left(B N\left(W_l^{(1)} \sigma\left(B N\left(x_l\right)\right)\right)\right)$                (14)

Such that $W_l^{(i)}$ is the $i$-th weight in the l-th module, and $i=1,2 . \sigma$ is the activation function (ReLU).

(2) Basic ResNet block

The ResNet module includes a basic block in addition to the identity mapping. If the input vector of ResNet in the l-th module is denoted as x_l, as illustrated in Figure 5(b), and the output vector of the l-th module is represented as x_(l+1) (which serves as the input for the next module), the output vector of the ResNet is consequently obtained.

The module is $x_{l+1}=\sigma\left[f\left(x_l,\left\{W_l^{(i)}\right\}\right)\right]+x_l$, where,

$f\left(x_l,\left\{W_l^{(i)}\right\}=\left(W_l^{(2)}\right)^T \sigma\left(\left(W_l^{(1)}\right)^T x_l\right)\right.$               (15)

(3) Basic ResNet block added BN

The l-th module with BN in ResNet is shown in Figure 5(c) and can be further described. The input feature of the l-th module is $x_l$, and the output of this layer is $x_{l+1}=f\left(x_l,\left\{W_l^{(i)}\right\}\right)+x_l$. In this case,

$f\left(x_l,\left\{W_l^{(i)}\right\}\right)=\left(W_l^{(2)}\right)^T \sigma\left(B N\left(\left(W_l^{(1)}\right)^T \sigma\left(B N\left(x_l\right)\right)\right)\right)$               (16)

where, BN is the batch normalization operation.

(4) Stability analysis of ResNet

The condition number

Lemma 1: $W_l^{(1)}, W_l^{(2)} \in \mathbb{R}^{n^2 \times n^2}$ are the two weight matrices of l-th module of ResNet. $I \in \mathbb{R}^{n^2 \times n^2}$ is the unit matrix, and $x_l \in \mathbb{R}^{n^2}$ is the input vector of l-th module. In this condition, the corresponding weight matrix of the whole module is $W_l^{(1)} W_l^{(2)}+I$, which satisfies the following inequality:

$\left(m_l^{(1)} m_l^{(2)}+1\right)\left\|x_l\right\|_2 \leq\left\|\left(W_l^{(1)} W_l^{(2)}+I\right)^T x_l\right\|_2 \leq\left(M_l^{(1)} M_l^{(2)}+1\right)\left\|x_l\right\|_2$             (17)

Such that $M_l^{(1)}, m_l^{(1)}$ are the maximum singular value and the minimum singular value of $W_l^{(1)}$, respectively, and $M_l^{(2)}, m_l^{(2)}$ are the maximum singular value and the minimum singular value of $W_l^{(1)}$, respectively.

Remark 1: According to inequality (1), the condition number of $W_l^{(1)} W_l^{(2)}+I$ can be defined as:

$\kappa_{l, R}\left(W_l^{(1)} W_l^{(2)}+I\right)=\frac{M_l^{(1)} M_l^{(2)}+1}{m_l^{(1)} m_l^{(2)}+1}$                 (18)

Remark 2: In this case, the weight matrix is always invertible. If the matrix is not full, the minimum singular value of the matrix becomes zero, and the condition number tends to infinity. Thus, opposing the actual classification network, the network diverges. Consequently, the matrix is full rank, and the singular values are non-negative.

Then, stability analysis of the ResNet block:

Based on the assumptions of Eq. (2), the condition number of $l$-th module of the base Network:

$\kappa_{l, B}\left(W_l^{(1)} W_l^{(2)}\right)=\frac{M_l^{(1)} M_l^{(2)}}{m_l^{(1)} m_l^{(2)}}$               (19)

Seemingly

$\kappa_{l, R}\left(W_l^{(1)} W_l^{(2)}+I\right)-\kappa_{l, B}\left(W_l^{(1)} W_l^{(2)}\right)<0$                  (20)

The statement above implies that the ResNet network has a lower condition number compared to the normal network assuming they have the same input features and convolutional kernel. As a result, the forward propagation convergent learning process is faster in ResNet compared to networks that are solely based on convolution. The evolutional property of ResNet can also be discussed when taking $l$-th module as an example. Let $\left(W_l+I\right)^T\left(x_l+\Delta x_l\right)=a_l+\Delta a_l$, where, $\Delta x_l$ is the absolute error of the input vector of the $l$-th module, $a_l$ designates the output vector of the $l$-th module operation, $\Delta a_l$ is the absolute error of the output vector after training. Then the perturbation estimates of the output data of $l$-th module can be represented in the following Lemma.

Lemma 2: The perturbation of the output data after the l-th module operation satisfies the following inequality:

$\frac{1}{\kappa_{l, R}\left(W_l+I\right)} \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|} \leq \frac{\left\|\Delta a_l\right\|}{\left\|a_l\right\|} \leq \kappa_{l, R}\left(W_l+I\right) \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|}$                 (21)

When the same idea provided in Lemma 2 is followed, the perturbation bounds about the basic block can be provided. Let $W_l^T\left(x_l+\Delta x_l\right)=b_l+\Delta b_l \cdot x_l, \Delta x_l$, which is the same as the case in ResNet. $b_l, \Delta b_l$ are the output vector and perturbation of the output, respectively. Thus, the relative error of the input data is limited to:

$\frac{1}{\kappa_{l, B}\left(W_l\right)} \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|} \leq \frac{\left\|\Delta b_l\right\|}{\left\|b_l\right\|} \leq \kappa_{l, B}\left(W_l\right) \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|}$               (22)

Meanwhile $\kappa_{l, R}\left(W_l^{(1)} W_l^{(2)}+I\right)-\kappa_{l, B}\left(W_l^{(1)} W_l^{(2)}\right)<0$, then the following can be deduced:

$\frac{1}{\kappa_{l, B}\left(W_l\right)} \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|} \leq \frac{1}{\kappa_{l, B}\left(W_l+\mathrm{I}\right)} \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|}$               (23)

$\kappa_{l, B}\left(W_l\right) \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|} \geq \kappa_{l, R}\left(W_l+I\right) \cdot \frac{\left\|\Delta x_l\right\|}{\left\|x_l\right\|}$                 (24)

In other words, from inequality (23) and inequality (24), it can be concluded that the upper limit of relative error of the output data will be smaller and the lower limit will be larger in ResNet CNN. The following Lemma can be used as evidence of the fact that ResNet with BN has better outcomes compared to the base network in divergence, such that ResNet with BN has a more stable learning process since it has a smaller perturbation range than the base network.

Let $x_{l+1}$ be the actual output of the $l$-th module, whereas $y_{l+1}$ is the actual output vector of the $l$-th module with BN. The loss function is $E=\frac{1}{2}\left(t_{l+1}-x_{l+1}\right)^2, E_{B N}=\frac{1}{2}\left(t_{l+1}-y_{l+1}\right)^2$. $\sigma$ designates the ReLU activation function, and $t_{l+1}$ designate the expected output vector of the $l$-th module.

Lemma 3: The output vectors of the ResNet module with BN and without BN can be written as $x_{l+1}=\left(W_l\right)^T \sigma\left(x_l\right)+x_l, y_{l+1}=\left(W_l\right)^T \sigma\left(B N\left(x_l\right)\right)+x_l$, respectively. The following equality can be achieved:

$\frac{\left\|\Delta B N\left(x_l\right)\right\|}{\left\|B N\left(x_l\right)\right\|} \leq \frac{\left\|\Delta x_l\right\|}{\|x l\|}$                 (25)

Table 2. Model training parameters

3.4 Model training

A total of 80 model iterations was carried out using a batch size of 16 and an initial learning rate of 0.001, all within a Colab environment. Table 2 provides a detailed breakdown of the training parameters.