Implemented classification techniques for osteoporosis using deep learning from the perspective of healthcare analytics

Abstract

BACKGROUND:

Osteoporosis is a medical disorder that causes bone tissue to deteriorate and lose density, increasing the risk of fractures. Applying Neural Networks (NN) to analyze medical imaging data and detect the presence or severity of osteoporosis in patients is known as osteoporosis classification using Deep Learning (DL) algorithms. DL algorithms can extract relevant information from bone images and discover intricate patterns that could indicate osteoporosis.

OBJECTIVE:

DCNN biases must be initialized carefully, much like their weights. Biases that are initialized incorrectly might affect the network’s learning dynamics and hinder the model’s ability to converge to an ideal solution. In this research, Deep Convolutional Neural Networks (DCNNs) are used, which have several benefits over conventional ML techniques for image processing.

METHOD:

One of the key benefits of DCNNs is the ability to automatically Feature Extraction (FE) from raw data. Feature learning is a time-consuming procedure in conventional ML algorithms. During the training phase of DCNNs, the network learns to recognize relevant characteristics straight from the data. The Squirrel Search Algorithm (SSA) makes use of a combination of Local Search (LS) and Random Search (RS) techniques that are inspired by the foraging habits of squirrels.

RESULTS:

The method made it possible to efficiently explore the search space to find prospective values while using promising areas to refine and improve the solutions. Effectively recognizing optimum or nearly optimal solutions depends on balancing exploration and exploitation. The weight in the DCNN is optimized with the help of SSA, which enhances the performance of the classification.

CONCLUSION:

The comparative analysis with state-of-the-art techniques shows that the proposed SSA-based DCNN is highly accurate, with 96.57% accuracy.

Keywords

Osteoporosis classification accuracy optimization bias neural network and squirrel search

1. Introduction

Healthcare analytics help identify patients at improved risk for specific disorders like osteoporosis or cancer by analyzing patient data from medical imaging, Electronic Health Records (EHRs), and clinical notes [1]. It enables healthcare professionals to act early, offer individualized treatment plans, and enhance results. Analytics are used to predict patient outcomes based on medical imaging data. Predictive models suggest insights into prognosis, illness development, and therapy response by analyzing models and stages in imaging data combined with other clinical factors. It helps healthcare providers make decisions about patient care [2].

Patient care, diagnosis, treatment planning, and medical research are all significantly impacted by healthcare analytics and image processing. Analyzing healthcare data to derive insights and guide well-informed decisions is known as healthcare analytics. It includes several methods, including ML, data mining, statistical analysis, and predictive modeling. Healthcare analytics may be used in medical imaging to extract useful information from massive volumes of image data, enabling improved diagnosis and treatment [3].

Biomedical imaging is crucial in assessing and treating osteoporosis, a disorder marked by reduced bone density and quality, making individuals more prone to fractures. Dual-Energy X-ray Absorptiometry (DEXA) is a primary imaging technique that employs low-dose X-rays to evaluate the density of minerals in bones, particularly in important regions such as the spine and hip. This assessment helps in diagnosing and determining the likelihood of fractures. Additional methods, such as Quantitative Computed Tomography (QCT) and Peripheral Quantitative Computed Tomography (pQCT), offer three-dimensional understanding of bone density, specifically in the trabecular and cortical sections. Magnetic Resonance Imaging (MRI) provides detailed information about the structure. In addition, ultrasound techniques such as Quantitative Ultrasound (QUS) offer a non-ionizing and portable method for evaluating peripheral bone density. These imaging techniques combined empower healthcare practitioners to diagnose osteoporosis, evaluate its extent, and track treatment outcomes, ultimately improving patient care and contributing to continuing research in bone health. Progress in biomedical imaging techniques is continuously enhancing our comprehension and treatment of osteoporosis, leading to better results for individuals impacted by this ailment.

Healthcare analytics can aid in optimizing resource allocation and operational effectiveness [4]. Hospitals may predict periods of high demand, distribute staff and resources accordingly, and improve the patient flow between departments, for instance, by analyzing imaging data and patient records. Image alteration, improvement, and analysis are all part of the Feature Extracting (FE) process from image data [5]. Image processing methods are used in the healthcare industry to improve medical images, extract characteristics, and aid in diagnosis and treatment planning.

Combining healthcare analytics with image processing enables more precise and practical analysis of medical images, improving patient results in diagnosis and treatment planning [6]. It equips healthcare providers with innovative tools and insights to make decisions based on the best available evidence and assign patients personalized treatment. Osteoporosis is a medical illness that causes bone tissue to deteriorate and lose density, which makes bones more brittle and increases the risk of fractures [7]. It is a chronic condition that often affects older people, especially women, after menopause, although it can also affect males and younger people. Osteoporosis is sometimes described as a “silent disease” since it frequently advances without warning until a fracture happens [8].

Osteoporosis frequently causes fractures in the hip, wrist, and spine (vertebral fractures). Fractures can cause discomfort, height loss, abnormalities, and restricted movement. By improving numerous elements of diagnosis, risk assessment, treatment planning, and research, computational algorithms have influenced the field of osteoporosis [9]. Computational algorithms have been created to determine a person’s risk of getting osteoporosis and are frequently integrated with Deep Learning (DL) methods. These algorithms estimate the chance of osteoporosis using a selection of data sources, including medical history, demographic data, lifestyle variables, and measures of bone density. These algorithms can help healthcare providers identify high-risk patients who may receive help from preventative measures or early intervention by analyzing massive datasets and finding pertinent risk variables [10].

When classifying osteoporosis, image processing may take bone density or texture measures from X-ray or DXA scans and feed them into a DL model. Individuals at risk for or diagnosed with osteoporosis should communicate with healthcare specialists for an accurate assessment, direction, and individualized management regimens. The effects of osteoporosis are lessened, and the risk of fractures is decreased with early identification, preventative measures, and suitable therapies [11].

There are several obstacles to overcome when applying DL to osteoporosis classification in medical analytics. The availability and quality of data provide difficulties; imbalances and incompleteness affect the generalization and performance of the model. The interpretability of DL models raises questions since it is challenging to comprehend decision-making processes, which are crucial to the healthcare industry, due to their opaque nature. It is difficult to generalise models across different patient groups, and ethical issues with patient privacy and data security must be considered. When it comes to DL models, there are problems with explainability and transparency, which makes it difficult to win over clinicians who require guidance on model predictions.

1.1 Research motivation

•
The research acknowledges the importance of bias initialization in Deep Convolutional Neural Networks (DCNNs) for effective learning dynamics and convergence, highlighting a crucial but often overlooked aspect in neural network applications.
•
DCNNs are employed for their capability to automate feature learning during the training phase, enabling the network to recognize relevant characteristics directly from the data without the need for manual feature engineering.
•
The research introduces the Squirrel Search Algorithm (SSA), a novel approach inspired by the foraging habits of squirrels. SSA combines Local Search (LS) and Random Search (RS) techniques, providing an efficient means of exploring the search space for optimal bias values in DCNNs.
•
SSA effectively balances exploration and exploitation in the search space, utilizing promising areas to refine and improve solutions. This approach is crucial for efficiently identifying optimum or nearly optimal solutions for bias initialization in DCNNs.

1.2 Research contributions

•
The research utilizes DCNNs to classify osteoporosis, capitalizing on their superior capabilities compared to conventional Machine Learning (ML) methods in the field of picture processing.
•
The research proposes the SSA, which draws inspiration from the foraging habits of squirrels, to enhance the optimization of bias values in the DCNN.
•
By enabling DCNNs to autonomously extract features during the training process, we have resolved the issue of time-consuming feature extraction in traditional ML algorithms.

The remainder of the article is organized as follows: An overview with recent research discussion is given in Section 1, related work is given in Section 2, the proposed Squirrel Search-based Deep Convolutional Neural Network (SSA-DCNN) is discussed in Section 3, the outcome of the SSA-DCNN with existing comparisons is illustrated in Section 4, and the article is concluded with future research recommendations in Section 5.
2. Related works

Several research studies have been conducted on osteoporosis, a complex illness with the clinical complication of bone fracture. Artificial Intelligence (AI) has made great strides in complicated data contexts where the human capacity to recognize high-dimensional correlations is restricted by recent developments in ML. Frequent consequences, including a substantial risk of fracture, are caused by diseases like osteoporosis, which significantly alter the structure of the trabecular bone. Five CNN models were used to analyze the hip radiographs and determine the degree of osteoporosis. The accuracy with which the CNN models identified osteoporosis from hip radiographs increased when clinical variables from patient records were included [12]. Visualizing the model performance overlapped the original image with a Gradient-based Class Activation Map (Grad-CAM). Additionally, we used 117 datasets for external validation. The proximal femur cortex and trabecular patterns, as well as internal and external validation sets, are accurately mapped by the outcomes of the Grad-CAM technique. One of the helpful screening techniques for simple osteoporosis prediction in a real-world clinical environment might be the DNN model [13]. A few DL research studies showed that integrating image features with patient features improved identification accuracy compared to image-only models. However, the extra impact of patient features on the image-only models has not been statistically recorded in earlier investigations.

A mandible’s textural and morphological features are analyzed using Dental Panoramic Radiographs (DPRs) to determine changes in bone density. In order to classify osteoporosis traits in DPRs, this study seeks to assess the discriminating ability of DCNNs used with various transfer learning algorithms. The problem of osteoporosis is common yet underdiagnosed. The work aimed to create a DCNN model to classify osteopenia and osteoporosis using lumbar spine X-ray images instead of dual-energy X-ray absorptiometry (DXA) measurements. Many researchers focus on neural networks and do not handle the complexity. To achieve efficient training, DCNNs need proper initialization of their weights [14]. Unsuitable initialization of weights can result in problems like disappearing or bursting gradients, which might obstruct convergence or result in instability during training. DCNN biases must be initialized carefully, much like their weights. Biases that are initialized incorrectly might affect the network’s learning dynamics and hinder the model’s ability to converge to an ideal solution. The weight optimization in DCNN uses the SSA.

The utilization of DL in radiographic pictures, specifically chest X-rays, is becoming increasingly significant due to the pressing requirement for precise and expeditious COVID-19 identification. DCNNs are frequently employed as reliable COVID-19 detectors, where Gradient Descent-Based (GDB) techniques are utilized to train their final fully-connected layers. Nevertheless, GDB algorithms encounter obstacles such as the need for human adjustment, vulnerability to local minima, and complications in parallelization using Graphics Processing Units (GPU). The Chimp Optimization Algorithm (ChOA) is suggested as a solution to these problems. It is designed to train the fully connected layers of DCNNs, with the goal of creating a rapid COVID-19 detector that can be implemented in parallel. The DCNN-ChOA is compared against conventional DCNNs, the Hybrid DCNN with Genetic Algorithm (DCNN-GA), and the Matched Subspace classifier with Adaptive Dictionaries (MSAD), utilizing datasets such as COVID-Xray-5k and COVIDetectioNet. The findings indicate that DCNN-ChOA achieves a detection accuracy of more than 99.11% while maintaining a false alarm rate of less than 0.89%, surpassing similar detectors in performance. The study employs Class Activation Maps (CAM) to detect regions affected by COVID-19, which are then compared to clinical results as confirmed by experts [15].

This work examines the application of DCNNs for identifying abnormal voices, specifically focusing on the difficulty of choosing the most effective DCNN structure. The Whale Optimization Algorithm (WOA) is utilized to autonomously select the optimal DCNN architecture, incorporating novel features such as a specialized encoding technique based on Internet Protocol Addresses (IPA) for simplified DCNN layer encoding and the advancement of variable-length DCNNs. The suggested model is assessed with pathological audio signals, attaining a maximum accuracy of 95.77%. It surpasses the second-best algorithm, VLNSGA-II, by 1.02% in terms of accuracy [16].

Furthermore, a study is dedicated to enhancing the efficiency of a DCNN architecture to achieve precise diagnosis of COVID-19 in lung CT scans. The Sine-Cosine Algorithm (SCA) is employed to optimize the structure of the DCNN. Enhancements include a novel encoding method based on internet protocol (IP) addresses and the addition of a weakened layer for variable-length DCNNs. The DCNN-IPSCA model surpasses both standard DCNNs and seven variable-length models in various performance metrics including sensitivity, accuracy, specificity, F1-score, precision, and receiver operative curve (ROC) and precision-recall curves. It achieves final accuracy rates of 98.32% and 98.01% on the SARS-CoV-2 and COVID-CT datasets, respectively [17].

A four-phase DL-based approach is suggested for real-time categorization in the context of underwater backscatter detection. A DCNN is utilized to extract features, while the Extreme Learning Machine (ELM) replaces the fully connected layer to enhance processing speed. The Hunger Games Search (HGS) tackles the uncertainty caused by ELM, while fuzzy systems maintain a balance between exploration and exploitation stages. The fuzzy HGS (FHGS) that was constructed has superior performance in detecting underwater anomaly targets, with an average improvement of 2.11% compared to the top-performing benchmark model [18].

Finally, a refined iteration of the ChOA is shown to autonomously discover the most effective DCNN structure for image classification tasks. When tested against 23 classifiers using nine benchmark image datasets, the proposed model demonstrates superior performance in the Fashion dataset and surpasses other benchmarks in 87 out of 95 evaluations. The utilization of the variable-length strategy, which involves the use of ChOA to independently develop DCNN architectures, is a significant contribution.

Trabecular bone, not compact cortical hard bone, is the main focus of osteoporosis detection to prevent bone fractures. In Table 1, system performance is compared to current trend techniques.

Table 1
The comparison of performance using various approaches

Model	Imaging modality	Classification module/ pre-trained networks	Validation	Accuracy (%)	AUC (%)
CNN [19]	X-ray	VGG16-TR-TF	5-fold cross	84.50	87.19
RNN [20]	MRI	LeNet- based CNN	3-fold cross	93.34	84.19
CNN [21]	X-ray	10-fold cross	10-fold cross	98.10	87.19
DCNN [22]	CT	BCC- net	23-fold cross	76.19	91.093
MultiTASK Scheme [23]	X-ray	MS-net	25-osteoporosis	92.38	90.20
ANN [24]	MRI	OSNet	200 cross	78.12	96.46
Feed-Forward Back Propagation [25]	X-RAY/DXA	ANN	Training-160 Testing-45	71.29	91.29
Multiple Regression Algorithm [26]	MRI	LR	30-fold	95.69	87.19
Features of BMD [27]	CT	ANN-LR	12-fold	96.19	90.10
ML [28]	MRI	SVM	15-fold	78.19	89.19
Deep Features [29]	MRI	RF	Training-120 Testing-75	88.19	93.01

Table 1 presents a comprehensive overview of different methods used for classifying imaging modalities. It highlights the utilization of convolutional neural networks (CNN), recurrent neural networks (RNN), artificial neural networks (ANN), and ML algorithms in various medical imaging modalities. The models utilize pre-trained networks, including VGG16-TR-TF, LeNet-based CNN, BCC-net, MS-net, OSNet, and SVM. Each of these networks is specifically designed for different imaging domains such as X-ray, MRI, and CT. Validation procedures vary greatly, ranging from 3-fold to 200 cross-validation schemes. Although the classification accuracies range from 71.29% to 98.1% and the area under the curve (AUC) values range from 84.19% to 96.46%, there is still a significant research gap. The lack of a defined validation technique and the unpredictability in selecting pre-trained networks impede direct comparisons of models. Moreover, the difference in the sizes of the datasets used for training and testing raises questions regarding the ability of the models to apply to new, unseen data. To address these deficiencies, it is recommended to implement consistent validation protocols, do comparative assessments of pre-trained networks, and investigate ensemble techniques. These actions can enhance the comprehension of ideal models for classifying imaging modalities, hence promoting progress in the field of medical image analysis [30].

3. SSA-DCNN

CNN shows impressive and revolutionary applications in different domains. CNNs have proven their tendency in modern medical image processing and are a powerful tool for manipulating complex image data. Recently, CNNs have discovered applications in different fields like sound examination, bioinformatics, Image Processing, and Medical Image Processing. Specifically, CNN appears vital in numerous Computer Vision (CV) tasks. This section shows the application of DCNN and its potential for efficiently classifying X-Ray images.

In the proposed work, an adapted CNN model for osteoporosis prediction has been designed and named an SSA-based DCNN model. The proposed SSA-based DCNN model comprises a set of multiple building blocks: five convolutional layers, five pool layers trailed by a pool level layer, a dropout layer, and three Fully Connected (FC) layers. The features evolving out of the FC layer are passed to the SoftMax Classifier, which classifies the osteoporosis image as healthy or defective. The architecture of the developed SSA-based DCNN model is shown in Fig. 1.

Figure 1.

Overall architecture of the SSA-based DCNN model.

The SSA-based DCNN model has two selections based on the chosen filter size (F). The proposed SSA-based DCNN (3 $\times$ 3) model comprises five convolutional layers, five pool layers trailed by a pool level layer, a drop-out layer, and three FC layers. The convolution layer uses a filter of size 3 $\times$ 3 at each level.

A ConvNet/CNN belongs to the DL domain with multiple variable factors for automated FE and learning. A minimal volume of pre-processing is sufficient for a CNN, and some applications even work well without any preprocessing. Compared to traditional techniques, CNN involves automated FE. The CNN works well for the image dataset due to image refinement and weight reusability in every iteration. The basic ConvNet is weighted with layers, wherein each layer transforms data using a differentiable function. The three base layers, including the design of ConvNet, are the convolutional, pooling, and FC layers. These layers are linked in a converted path to shape the ConvNet model.

CNN processes the raw pixel data by traversing through the layers compiled in the network. The convolution layer applies filters to the input data to efficiently capture the spatial and temporal features of the data. Pooling layers perform dimensionality reduction, thereby decreasing the computational complexity without compromising the critical features suitable for the application. The pooling layer performs a maximum or Average operation on the image section overlapped by the kernel. Max pooling is a standard operation applied in various applications. The FC layer relates to the data evolving out of convolutional layers, and the last layer corresponds to the SoftMax Classifier that performs the classification task.

3.1 Convolution layer

The convolutional layer is a design that looks like the workings of the human cortex. The neurons related to the convolutional layer are associated with neighborhood areas of the input, and the outputs generated depend on those nearby regions. This layer is defined by a group of learnable image filters called kernels or Feature locators. Convolution is an essential linear operation where a kernel is convolved with an osteoporosis image, an array of integers called tensors. An elementwise multiplication is performed between each kernel element and the osteoporosis input, forming a summation that constitutes the output, including the osteoporosis feature map. This process is repeated throughout the image, creating various feature maps representing the input’s osteoporosis characteristics. The color osteoporosis image is resized to 224 $\times$ 224 and is passed as input to the SSA-based DCNN model. Consider ( $W_{i},H_{i},D_{i}$ ) the osteoporosis input volume’s width, height, and depth, where $D_{i}$ refers to the number of channels in the osteoporosis image. As this model does not need any preprocessing and runs on the raw data, which includes all the three-color channels, namely {Red, Blue, and Green}, the value of $D_{i}$ is fixed at 3. Then, the output volume $W_{0}\times H_{0}\times D_{0}$ is determined by applying Eq. (1).

$\displaystyle W_{o}=\frac{w_{i}-F+2P}{S}+1$ (1)

(Width ( $W_{0}$ ), Height ( $H_{0}$ )) are computed equally by symmetry and are given by Eq. (2).

$\displaystyle H_{o}=\frac{H_{i}-F+2P}{S}+1$ (2)

where $F\to$ filter size, $P\to$ padding quantity, and $S\to$ stride parameter.

Stride refers to the number of pixels the filter moves across the osteoporosis image after every convolution function. The same padding is applied to allow the center of the kernel to intersect with the outermost element of the input data. It ensures that the filter is applied to all the elements of the osteoporosis input. F and S are the hyperparameters in this layer set to 3 and 1, respectively. Filter moves pixel by pixel with stride set to 1.

The 3 $\times$ 3 filter is assigned with random values using normal distribution and is convolved with a 3 $\times$ 3 window in the osteoporosis input patch highlighted as a red rectangle. The dot product between the filter and osteoporosis 3 $\times$ 3 window is performed, and the resulting summation value (147 in this case) is placed in the osteoporosis feature map highlighted in red. This convolution process is frequently requested on the entire osteoporosis image patch by moving the filter one pixel at a time, and the osteoporosis feature map is produced for the given osteoporosis patch. The first convolution layer performs this convolution operation with 32 filters, and the output produced is subjected to the Rectified Linear Unit (ReLU) Activation Function.

3.2 Activation function

The Activation Function and its placement in the CNN model are crucial for achieving superior results. Analyses complex connections between osteoporosis attributes in a network. It identifies the data that should be sent to the network’s end and not to the end. Some Activation Functions popularly used are ReLU, SoftMax, tanh, and Sigmoid.

The proposed model, namely the SSA-based DCNN (3 $\times$ 3) model, uses the ReLU Activation Function. Rectified Linear Unit (ReLU) is the common Activation Function applied in Deep Neural Networks. ReLU is used to introduce non-linearity in the osteoporosis image. Except for the last layer, ReLU is applied as an Activation Function for the remaining layers [Vinod et al. 2010]. The computation of ReLU $\emptyset\left(y\right)$ is given according to Eq. (3).

$\displaystyle\emptyset\left(y\right)\to\max\left({0,y}\right)$ (3)

were, $y\to$ osteoporosis feature element.

This procedure of convolution and ReLU is repeatedly applied to the entire osteoporosis image of size 224 $\times$ 224 for all three image channels by sliding the filter one pixel at a time. The final osteoporosis feature map generated from the first convolution layer (Conv 1) is sent to the pooling layer (POOL1) for dimensionality reduction. The entire osteoporosis image of size 224 $\times$ 224 $\times$ 3 is processed using Eqs (1) and (2), which produce an osteoporosis feature map and are sent to the Pool 1 layer for further processing. The number of learnable parameters $P_{c}$ in the convolution layer is given as per Eq. (4).

$\displaystyle P_{c}=\left({\left({m\times n\times F_{p}}\right)+1}\right)% \times F_{C}$ (4)

where $m\to$ is the shape of the width of the filter, $n\to$ is the shape of the height filter, $F_{p}\to$ is the number of filters in the previous layer, and $F_{c}\to$ is the number of filters.

The number of parameters in CONV1 is computed as 896 using Eq. (4).

3.3 Pooling layer

The pooling layer is associated with the convolutional layer to downsample the feature maps without losing the principal osteoporosis features. Consequently, it reduces the number of edges to learn and the number of calculations needed in the network. The pooling layer computes the features in the feature map designed by the convolution layer. Thus, further tasks are performed to deal with osteoporosis rather than aggregate osteoporosis features produced by the convolution layer. This simplified processing covers the features needed for the model to achieve robustness in coping with the different osteoporosis features of the osteoporosis image. Pooling is accomplished by sliding a two-dimensional filter over each convolutional layer’s feature map channel and then processing the resulting features in the filtered region.

The feature map arising from the convolution layer is subsampled with R $\times$ R non-overlapped regions (windows). This window is convolved with the feature map to change the input volume using the Max operation spatially. Each scan of this window infers the maximum activation value, indicating the most significant osteoporosis feature, while other unique features are forgotten. The thought behind Max pooling is to appropriately reduce the feature map for added computation, reducing the resulting complexity in the network. The Max pooling operation picks the maximum value from the target window. It is possible that the model fits well for the training data, and hence, it becomes difficult to generalize the model, which leads to overfitting. Overfitting in CNN can be prevented by using a regularization method called dropout. The flattened output of the 11358-osteoporosis feature vector is sent to the drop-out layer for further computation.

Considering the input volume size as $W_{i}\times H_{i}\times D_{i}$ and Max pooling window size as R $\times$ R, the DR feature element included in the output volume $W_{0}\times H_{0}\times D_{0}$ is determined by using the Eq. (5).

$\displaystyle w_{o}=\frac{w_{i}-R}{S}+1$ (5)

Where $S\to$ S tride parameter ( $W_{o}$ , $H_{o}$ ) are computed equally by symmetry and is given by Eq. (6).

$\displaystyle H_{o}=\frac{H_{i}-R}{S}+1$ (6)

were, $F\to$ window size; $S\to$ stride parameter.

3.4 Dropout layer

The flattened output osteoporosis feature vector of size 11358 is received at the drop-out layer. Dropout is a regularization technique to avoid data overfitting in the network, which is applied in the training phase of the model. Overfitting is an issue where the model does not learn or memorize the osteoporosis training data well. In cases of overfitting, multiple neurons will extract similar osteoporosis features from the osteoporosis image. Hence, neurons in the layer are randomly set to zero as a dropout factor. The fraction of dropped-out neurons is termed the dropout rate. Based on the dropout rate, a different network is created at each training step as it involves the arbitrary rejection of specific neurons.

A network with a dropout layer is less prone to overfitting as it gives better generalization. The network can learn robust osteoporosis R features thanks to this generalization approach. A dropout rate 0.5 is set, indicating the probability of neurons’ retention in the network. Different networks are created based on the dropout rate, and the process of repeatedly applying convolution and pooling is experienced for 100 iterations. The resulting osteoporosis feature vector is passed to the fully connected layer for computation.

3.5 Fully connected layer

A Deep Neural Network winds up with an FC layer. The FC layer interconnects its associated nodes with the nodes of the successive FC layers. These layers are, in general, those of neurons. Usually located before the output layer, the FC layer connects neurons between two distinct layers. It modifies the osteoporosis image from prior layers and communicates it to the FC layer. The osteoporosis feature vector of size 11358 was generated as output mapped onto 2048 neurons at the initial FC layer FC1, trailed by ReLU activation. The subsequent connected layer, FC2, maps 2048 neurons to 256 neurons and applies ReLU activation for output generation. The final FC layer, FC3, converges to 64 from 256 neurons, trailed by ReLU activation. The resulting output from the connected layers corresponds to 64 neurons mapped to the value 2, indicating the number of classes needed for osteoporosis Classification. This osteoporosis $`R^{\prime}$ output vector is passed to the SoftMax Classifier for osteoporosis Classification.

The number of parameters of the FC layer to the convolution layer is labeled as $P_{cf}$ and is given as per Eqs (7) to (9).

$\displaystyle P_{cf}=W_{cf}+B_{cf}$ (7) $\displaystyle W_{cf}=O^{2}\times N\times F$ (8) $\displaystyle B_{cf}=F$ (9)

Where $O\to$ output of previous convolution layer; $N\to$ number of kernels in the last layer convolution, $F\to$ Number of neurons in the FC layer, $W_{cf}\to$ weight factor between FC layer to convolution layer, and CcfBias factor between FC layer to convolution layer. The number of parameters between FC layers labeled as $P_{ff1}$ is given per Eq. (10) to Eq. (12).

$\displaystyle P_{\textit{ff1}}=W_{\textit{ff1}}+B_{\textit{ff1}}$ (10) $\displaystyle W_{\textit{ff1}}=F_{p}\ast F$ (11) $\displaystyle B_{\textit{ff1}}=F$ (12)

$F_{p}$ is the number of neurons in the previous layer, F is the number of neurons in the current layer, $W_{\textit{ff1}}\to$ is the weight factor between FC layers, and $B_{\textit{ff1}}$ is the bias factor between FC layers.

3.6 SoftMax classifier

The osteoporosis output vector from the FC3 layer is fed to the SoftMax Classifier. SoftMax activation is applied to the output vector to generate the likelihood values for the osteoporosis Classification. The SoftMax Classifier (SMC) is a directed model exploited for training that normalizes the scores acquired into probabilities and performs binary osteoporosis classification. The variables of SMC are configured in such a path as to limit the cost function while executing the gradient descent algorithm. It allows the model to be trained to extract the osteoporosis discriminative features from the osteoporosis image. Given the osteoporosis vector ‘ $y$ ’, SoftMax computation $S(y_{i})$ is described according to Eq. (13).

$\displaystyle S\left({y_{i}}\right)=\frac{e^{y_{i}}}{\mathop{\sum}\nolimits_{j% =0}e^{y_{i}}}$ (13)

Where ‘ $y_{i}$ ’ corresponds to osteoporosis vector values. These osteoporosis vector values are turned into the probability of each osteoporosis class of ‘R’. The target during learning at every iteration is to limit the Mean Squared Error (MSE), provided according to Eq. (14). The gradient descent algorithm prepares the model with the gradient. The assessed predictions are analyzed against the ground truth target iteratively until the cost value reaches a minimal weight.

$\displaystyle\textit{MSE}=\frac{1}{n}\left({y_{i}-y_{i}^{\prime}}\right)^{2}$ (14)

Where, $n\to$ number of classes, $y_{i}\to$ actual output, and $y_{i}^{\prime}\to$ observed output.

Adjusting the weights during backpropagation results in learning. The gradient descent algorithm is applied during backpropagation to adjust the weights in the kernel. The updated weight $w_{\textit{new}}$ is computed.

$\displaystyle w_{\textit{new}}=w_{\textit{old}}-\alpha\frac{\partial\left(% \textit{MSE}\right)}{\partial\left(w\right)}$ (15)

where, $w_{\textit{old}}\to$ corresponds to the old weight value; $\alpha\to$ corresponds to the learning rate, set to 0.0005 and $\partial\to$ the differentiation function.

Table 2

Configurations of the SSA-based DCNN (3 $\times$ 3) model

Layer	Kernel shape	Kernel	Stride	Learnable parameters
Conv 1	224 $\times$ 224 $\times$ 3	32	1	896
Pool 1	112 $\times$ 112 $\times$ 3	32	2	0
Conv 2	112 $\times$ 112 $\times$ 3	64	1	18496
Pool 2	56 $\times$ 56 $\times$ 3	64	2	0
Conv 3	56 $\times$ 56 $\times$ 3	128	1	73856
Pool 3	28 $\times$ 28 $\times$ 3	128	2	0
Conv 4	28 $\times$ 28 $\times$ 3	256	1	295168
Pool 4	14 $\times$ 14 $\times$ 3	256	2	0
Conv 5	14 $\times$ 14 $\times$ 3	256	1	590080
Pool 5	7 $\times$ 7 $\times$ 3	256	2	0
FC 1				25692160
FC2				524544
FC3				16448
SoftMax				258
Parameters				27211906

The kernel values are adjusted based on Eq. (15) to reduce MSE. In this work, this process of weight updation is repeated for 100 iterations.

Number of parameters in SoftMax Layer ‘ $S_{p}$ ’.

$\displaystyle S_{p}=\left({\left({F_{p}\ast F}\right)+}\right)\ast F$ (16)

where, $F_{p}\to$ Number of neurons in previous layer; $F\to$ Number of neurons in current layer, $S_{p}$ is computed as 258 using Eq. (16). The convolution and Pooling configuration of the SSA-BASED DCNN (3 $\times$ 3) model is summarized in Table 2.

The number of learnable parameters in the network increases based on the filter size chosen in the convolution layer. This work, a 3 $\times$ 3 filter is used, and their linked learnable parameters are performed in Table 2. From Table 1, the SSA-Based DCNN model with a 3 $\times$ 3 filter has fewer parameters to be learned, equal to 27211906. The smaller the filter size, the fewer the parameters to be learned.

Training data is processed per Algorithm 1 to perform osteoporosis prediction for each population with a bias and a weight value. Each training data point is considered a squirrel in this work. Squirrel location is updated on each iteration based on CA, which is regarded as the fitness value. In this algorithm, CA is viewed as a fitness value. A better population lying in that direction is selected as the optimal weight and bias values based on the parameters, namely movement, speed, and angle of action. The algorithm’s execution results in optimal weight and bias values that best fit the osteoporosis training data used in this work. These selected values are selected as input to the Squirrel search-based DCNN classifier, which performs osteoporosis classification.

Algorithm 1 for Improved Squirrel Search Algorithm
Input: Training Data $\to$ TrData; Training Label $\to$ TrLabel; Testing Data $\to$ TeData; Testing Label $\to$ TeLabel.
Output: Optimal values of weight and bias ( $w, b$ )
Step 1. Begin
Step 2. Randomly Initialize the population
Step 3. For Everyone in the population
Step 4. Calculate the fitness value as Classification Accuracy (CA)
Step 5. Update the population based on the $G_{\textit{best}}$ value.
Step 6. End For
Step 7. If all individual’s position is not updated
Step 8. Go To Step 2
Step 9. Else
Step 10. Find the best parameter values based on the $G_{\textit{best}}$ value.
Step 11. Find the seasonal transitional judgment and update the position of the individual.
Step 12. End If
Step 13. Repeat till convergence is achieved.
Step 14. End

In this work, the SSA is included with the DCNN classifier for accurate osteoporosis Classification. SSA is applied for optimal bias and weight value selection, integrated with the Support Vector Machine (SVM) to ensure accurate osteoporosis recognition. The osteoporosis feature is fed to the Squirrel search-based DCNN classifier, and its performance is assessed. Algorithm 2 shows the steps of an SSA-based DCNN classifier to train the osteoporosis features. Based on the feature vector, the trained SSA-based DCNN classifier is then used to classify whether the given input osteoporosis image is healthy or defective.

Algorithm 2 for SSA-based DCNN Classifier
Input: Reduced osteoporosis feature set from ASO algorithm (RFV), Optimal parameters from improved squirrel search algorithm ( $w, b$ )
Output: Healthy or osteoporosis defective
Step 1. Each ( $x_{i}$ ) belongs to X Do // $x_{i}$ , a vector containing osteoporosis features.
Step 2. Classify ( $x_{i}$ ) using $f(x_{i})$
Step 3. If there is a misclassification
Step 4. Perform gradient update using regularization parameters and loss.
Step 5. Else
Step 6. Perform gradient update using regularization parameter.
Step 7. Repeat
Step 8. End

4. Results and discussion

The proposed work has been conducted using Python, and its performance has been assessed for the osteoporosis dataset. The performance of the proposed model has been evaluated using the metrics of Accuracy, Precision, MSE, and Recall. The performance of the proposed approach is compared with the existing methods, namely gradient-based class activation map (Grad-CAM), CNN, and DCNN. Accuracy, Precision, and Recall results are given as per Eqs (17) to EQU(19). The images used in this research are presented in Figs 2 and 3.

Figure 2.

A flowchart depicting the process of selecting patients.

This study was a retrospective, multicenter study in a Chinese hospital with four sub-districts and another sizeable tertiary centre. The hospital’s institutional review board and ethics committee waived informed consent for the study. Medical and imaging data were collected from January to May 2023. A total of 6246 images were examined in the final analysis of the training and testing model. These images had a mean age of 60 years $\pm$ 9.3 years. Table 3 details clinical and demographic parameters for training, validation, and two test cohorts, showing no significant differences in gender, age, or BMI in these groups.

$\displaystyle\textit{Accuracy}=\frac{\textit{TP}+\textit{TN}}{\textit{TP}+% \textit{TN}+\textit{FP}+\textit{FN}}$ (17) $\displaystyle\textit{Precision}=\frac{\textit{TP}}{\textit{TP}+\textit{FP}}$ (18) $\displaystyle\textit{MSE}=\frac{1}{n}\mathop{\sum}\limits_{i=1}^{n}\left({Y_{i% }-\widehat{Y_{i}}}\right)^{2}$ (19)

True positives (TP) indicate the number of osteoporosis-defective images classified as osteoporosis-defective, and true negatives (TN) show the number of healthy images classified as healthy. False positives (FP) indicate the number of healthy images classified as DR defective. False negatives (FN) indicate the number of osteoporosis-defective images that are classified as healthy. ‘ $n$ ’ is a data point, and experimental value is implied by $Y_{i}$ and predicted value is indicated by $\widehat{Y_{i}}$ .

Table 3

Features of 5614 participants’ demographics

Attributes	Training size	Validation size	Test 1 size	Test 2 size	Total size
Patients	5614	632	632	632	7510
Age	60	61	60.5	61.5	60.6
Male	2526	201	198	195	3120
Female	3088	431	434	437	4390
BMI (kg.m²)	24.17	24.16	24.07	23.98	23.19
T-score	$-$ 1.89	$-$ 1.90	$-$ 1.81	$-$ 1.82	$-$ 1.92
BMI (%)
Osteopenia	1998	226	228	230	231
Osteoporosis	1768	225	227	226	229

Figure 3.

Sample Images.

According to Eq. (17), accuracy serves as a proxy for the overall accuracy of the osteoporosis fundus image classification. Precision is the measure of correctly identifying an osteoporosis image out of the total number of classified positive images given as per Eq. (18). The average squared difference between the predicted and actual values in a regression issue is frequently measured using the MSE metric. It measures the model’s precision by quantifying the average squared difference between EQU’s predicted and true values Eq. (19). The confusion matrix obtained for the various datasets using the proposed SSA-based DCNN model for two different filter sizes 3 $\times$ 3 during a particular run is shown in Tables 4 to 6.

Table 4

Comparison of accuracy

Iteration	Grad-CAM	CNN	DNN	SSA-based DCNN
100	89	91	92	93
200	90	91.5	92.5	94
300	91	93	94	95
400	92	93.5	94.5	96.57

In terms of overall performance, the SSA-Based DCNN model consistently outperforms the other models, achieving the highest accuracy of 96% at 400 iterations. The CNN and DNN models show comparable performance, with the DNN model slightly outperforming the CNN model. The Grad-CAM model lags the other models’ accuracy but improves over iterations.

However, precision and other metrics can improve model performance analysis, so accuracy alone may not be enough. To conduct more comprehensive research, it would be beneficial to have more information, such as TP and FP, to determine precision and assess the models’ performance more accurately. The SSA-based DCNN model starts with an accuracy of 93% and consistently improves, reaching 97% at the end of 400 iterations.

Figure 4.

Comparison of accuracy.

Figure 5.

Comparison of precision.

Figure 6.

Comparison of MSE.

Based on these observations, it is tentatively identified that SSA-based DCNN consistently performs better than DNN and CNN in terms of MSE.

4.1 Discussion

The comparative analysis in Table 4 reveals the performance dynamics of different models – Grad-CAM, CNN, DNN, and SSA-Based DCNN – across multiple iterations, focusing on accuracy as the primary metric. Strikingly, the SSA-based DCNN consistently outshines its counterparts, achieving the highest accuracy of 96.57% at 400 iterations. CNN and DNN exhibit comparable performances, with the DNN model slightly edging ahead. Although initially lagging in accuracy, Grad-CAM demonstrates a noteworthy improvement over successive iterations. However, accuracy alone may not provide a comprehensive understanding of model performance.

Table 5
Comparison of precision

Iteration	Grad-CAM	CNN	DNN	SSA-based DCNN
100	87	89	92	93
200	88	90	93	95
300	89	90.5	94	96
400	90	91	95	97

Table 6

Comparison of MSE

Iteration	Grad-CAM	CNN	DNN	SSA-based DCNN
100	691	625	587	413
200	678	598	523	389
300	625	542	498	356
400	567	512	456	323

To delve deeper into the evaluation, precision is explored in Fig. 5 and detailed in Table 5. Precision provides insights into the models’ ability to identify positive instances accurately. The SSA-based DCNN consistently demonstrates superior accuracy, reaching 97% at 400 iterations. This highlights the model’s effectiveness in correctly classifying positive cases, a crucial aspect of medical applications.

Table 7

Proposed SSA-based DCNN model prediction training and testing Confusion Matrix (CM)

		Osteoporosis	Normal	Osteoporosis	Normal	Osteoporosis	Normal
Truth	Osteoporosis	198	0	185	0	178	1
	Normal	4	81	67	28	68	119

Figure 7.

Proposed SSA-based DCNN model prediction training and testing Confusion Matrix (CM).

Furthermore, the analysis extends to Mean Squared Error (MSE), showcased in Table 6, which illuminates how well the models’ predictions align with the ground truth. The SSA-Based DCNN consistently outperforms CNN and DNN, indicating its proficiency in minimizing prediction errors. The graphical representations in Figs 4 to 6 further underscore the trends observed, providing a visual narrative of the models’ performance. In conclusion, the SSA-Based DCNN achieves higher accuracy and exhibits superior precision and reduced prediction errors compared to Grad-CAM, CNN, and DNN, offering a comprehensive and nuanced perspective on its excellence in osteoporosis detection in cranial bone scans.

The proposed SSA-Based DCNN Model’s classification CM is displayed in Table 7. These matrices report the total TP, FP, TN, and FN results (Fig. 7).

5. Conclusion and future work

In this article, the convolution and Max pool layers were put together in order to make a DL system called the SSA-based DCNN model. The ReLU activation unit was used to present non-linearity between the layers. A dropout layer integrated with the network prevented the network from memorizing the input, avoiding overfitting the osteoporosis information. The final feature vector generated was authorized for the SoftMax classifier that enables osteoporosis classification. The SSA-based DCNN model has been assessed on osteoporosis images from benchmark datasets. The network has been configured and assessed using filters of size 3 $\times$ 3, where the proposed approach achieves a minimal MSE rate of 323 and the highest accuracy of 96.57%. Using DL in healthcare analytics for osteoporosis classification highlights issues with model interpretability, generalization across different patient groups, and dataset quality. It is crucial to address ethical problems, such as data privacy and the requirement for precise, understandable models. Deployment is challenging since validation, appropriate measurements, and smooth integration into clinical workflows are essential. The resource-intensive nature and long training times of DL models, alongside continuous updates for evolving healthcare protocols, underscore the multifaceted nature of this approach. The primary contributions of the area promise improved diagnostic accuracy for prompt clinical interventions by taking a collaborative, multidisciplinary approach to these problems.

Many researchers do not widely examine the osteoporosis classification, and the research will be extended for segmentation and feature learning techniques. Future work should explore the robustness of the SSA-based DCNN model in diverse clinical scenarios and datasets. Investigating the model’s performance across different imaging modalities and patient demographics can enhance its generalizability. Additionally, refining the interpretability of the SSA algorithm and incorporating explainability features into the model architecture would contribute to building trust in its diagnostic decisions. Integrating real-world clinical feedback and validation in larger-scale studies would further validate the model’s efficacy in practical healthcare settings.

Footnotes

Conflict of interest

None to report.

Funding

None to report.

References

Mehta

Pandit

Shukla

. Transforming healthcare with big data analytics and artificial intelligence: A systematic mapping study. Journal of Biomedical Informatics. 2019; 100: 103311.

Banerjee

Chakraborty

Kumar

Biswas

. Emerging trends in IoT and big data analytics for biomedical and health care technologies. Handbook of Data Science Approaches for Biomedical Engineering. 2020; 121-152.

Panayides

Amini

Filipovic

Sharma

Tsaftaris

Young

\ldots

Pattichis

. AI in medical imaging informatics: current challenges and future directions. IEEE Journal of Biomedical and Health Informatics. 2020; 24(7): 1837-1857.

Panayides

Amini

Filipovic

Sharma

Tsaftaris

Young

\ldots

Pattichis

. AI in medical imaging informatics: current challenges and future directions. IEEE Journal of Biomedical and Health Informatics. 2020; 24(7): 1837-1857.

Zhang

Ning

Wang

Dong

Liu

\ldots

Zhang

. Deep learning of lumbar spine X-ray for osteopenia and osteoporosis screening: A multicenter retrospective cohort study. Bone. 2020; 140: 115561.

Lee

Jung

Ryu

Shin

Choi

. Evaluation of transfer learning with deep convolutional neural networks for screening osteoporosis in dental panoramic radiographs. Journal of Clinical Medicine. 2020; 9(2): 392.

Yamamoto

Sukegawa

Yamashita

Manabe

Nakano

Takabatake

\ldots

Yorifuji

. Effect of patient clinical variables in osteoporosis classification using hip x-rays in deep learning analysis. Medicina. 2021; 57(8): 846.

Jang

Choi

Kim

Chang

Yoon

Kim

. Prediction of osteoporosis from simple hip radiography using deep learning algorithm. Scientific Reports. 2021; 11(1): 19997.

Yamamoto

Sukegawa

Kitamura

Goto

Noda

Nakano

\ldots

Ozaki

. Deep learning for osteoporosis classification using hip radiographs and patient clinical covariates. Biomolecules. 2020; 10(11): 1534.

10.

Yamamoto

Sukegawa

Kitamura

Goto

Noda

Nakano

\ldots

Ozaki

. Deep learning for osteoporosis classification using hip radiographs and patient clinical covariates. Biomolecules. 2020; 10(11): 1534.

11.

Sukegawa

Fujimura

Taguchi

Yamamoto

Kitamura

Goto

\ldots

Furuki

. Identification of osteoporosis using ensemble deep learning model with panoramic radiographs and clinical covariates. Scientific Reports. 2022; 12(1): 1-10.

12.

Zhuang

Jiang

Xiangjie

Kong

. Progressive Distributed and Parallel Similarity Retrieval of Large CT Image Sequences in Mobile Telemedicine Networks. Wireless Communications and Mobile Computing, 2022; doi: 10.1155/2022/6458350.

13.

Zhang

Zhai

Cheng

\ldots

Wang

. Unraveling of Advances in 3D-Printed Polymer-Based Bone Scaffolds. Polymers. 2022; 14(3). doi: 10.3390/polym14030566.

14.

Yang

Xiao

Liu

Yin

\ldots

Zheng

. Iterative reconstruction of low-dose CT based on differential sparse. Biomedical Signal Processing and Control. 2023; 79: 104204. doi: 10.1016/j.bspc.2022.104204.

15.

Liu

Hou

Yang

Liu

Yin

\ldots

Zheng

. Soft Tissue Feature Tracking Based on Deep Matching Network. Computer Modeling in Engineering & Sciences. 2023; 136(1): 363-379. doi: 10.32604/cmes.2023.025217.

16.

Sun

Zhang

Wang

Tiwari

. Few-Shot Class-Incremental Learning for Medical Time Series Classification. IEEE Journal of Biomedical and Health Informatics. 2023. doi: 10.1109/JBHI.2023.3247861.

17.

Yang

Bo,

. Analysis and Design of Surgical Instrument Localization Algorithm. Computer Modeling in Engineering & Sciences. 2023; 137(1): 669-685. doi: 10.32604/cmes.2023.027417.

18.

Wang

Wipf

Cai

\ldots

. Sparse Bayesian Learning for End-to-End EEG Decoding. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2023; 45(2): 15632-15649. doi: 10.1109/TPAMI.2023.3299568.

19.

Wang

Verbeek

Schultz

\ldots

Liu

. Outlier detection using iterative adaptive mini-minimum spanning tree generation with applications on medical data. Frontiers in Physiology. 2023; 14. doi: 10.3389/fphys.2023.1233341.

20.

Gao

Pan

Shao

Jiang

Jin

\ldots

. Automatic interpretation and clinical evaluation for fundus fluorescein angiography images of diabetic retinopathy patients by deep learning. British Journal of Ophthalmology. 2022; 2022-321472. doi: 10.1136/bjo-2022-321472.

21.

Song

Xie

Zhu

Pan

Zhou

. Effects of PEMFs on Osx, Ocn, TRAP, and CTSK gene expression in postmenopausal osteoporosis model mice. International Journal of Clinical and Experimental Pathology. 2018; 11(3): 1784-1790.

22.

Wang

Gao

Zhong

\ldots

Lei

. A New Dressing System Reduces the Number of Dressing Changes in the Primary Total Knee Arthroplasty: A Randomized Controlled Trial. Frontiers in Surgery. 2022; 9. doi: 10.3389/fsurg.2022.800850.

23.

Chen

Zhi

Ding

Huang

Wang

Zheng

Wang

Miao

. lncRNA Xist Regulates Osteoblast Differentiation by Sponging miR-19a-3p in Aging-induced Osteoporosis. Aging and Disease. 2020; 11(5): 1058-1068. doi: 10.14336/AD.2019.0724.

24.

https://www.kaggle.com/code/favadhassanjaskani/osteoporosis-detection-in-bone-using-cnn/input.

25.

Cai

Gou

Khishe

Mohammadi

Rashidi

Moradpour

Mirjalili

. Improved deep convolutional neural networks using chimp optimization algorithm for Covid19 diagnosis from the X-ray images. Expert Systems with Applications. 2023; 213: 119206.

26.

Khishe

. An automatic COVID-19 diagnosis from chest X-ray images using a deep trigonometric convolutional neural network. The Imaging Science Journal. 2023; 71(2): 128-141.

27.

Yao

Chi

Khishe

. Parkinson’s disease and cleft lip and palate of pathological speech diagnosis using deep convolutional neural networks evolved by IPWOA. Applied Acoustics. 2022; 199: 109003.

28.

Martín

Khishe

Boostani

. COVID-19 diagnosis using chest CT scans and deep convolutional neural networks evolved by IP-based sine-cosine algorithm. Medical & Biological Engineering & Computing. 2022; 60(10): 2931-2949.

29.

Khishe

Mohammadi

Ramezani Varkani

. Underwater backscatter recognition using deep fuzzy extreme convolutional neural network optimized via hunger games search. Neural Processing Letters. 2023; 55(4): 4843-4870.

30.

Khishe

Azar

Hashemzadeh

. Variable-length CNNs evolved by digitized chimp optimization algorithm for deep learning applications. Multimedia Tools and Applications. 2023; 1-19.

Iteration	Grad-CAM	CNN	DNN	SSA-based DCNN
100	691	625	587	413
200	678	598	523	389
300	625	542	498	356
400	567	512	456	323

Iteration	Grad-CAM	CNN	DNN	SSA-based DCNN
100	691	625	587	413
200	678	598	523	389
300	625	542	498	356
400	567	512	456	323

Implemented classification techniques for osteoporosis using deep learning from the perspective of healthcare analytics

Abstract

BACKGROUND:

OBJECTIVE:

METHOD:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

1.1 Research motivation

Table 1 The comparison of performance using various approaches

3.5 Fully connected layer

Table 5 Comparison of precision

Footnotes

Conflict of interest

Funding

References

Table 1
The comparison of performance using various approaches

Table 5
Comparison of precision

Iteration	Grad-CAM	CNN	DNN	SSA-based DCNN
100	691	625	587	413
200	678	598	523	389
300	625	542	498	356
400	567	512	456	323