Lung Diseases Diagnosis-Based Deep Learning Methods: A Review

Abstract


Introduction
In affluent countries, the superiority of the spread of chronic disease has increased continuously because of prominent risk factors, according to the 2017 GBD.Coronavirus Disease 2019 (COVID-19), Pneumonia, and tuberculosis (TB) are examples of lung diseases.Every year, an estimated 334 million cases of asthma lead to fatalities, while TB results in the death of 1.4 million people, and 1.6 million people die from lung cancer.Additionally, pneumonia claims the lives of millions of people.According to the Forum of International Respiratory Societies [1], there is a high healthcare burden, and many people are infected [2].Fig. 1 shows the statistics of the world health organization (WHO) for lung diseases of the respiratory system [3].In today's world and according to the tremendous rate of death causes and disability, it can be shown clearly how lung disease is a leading cause in those statistics.Detecting a disease in its early stages can significantly increase the likelihood of recovery and long-term survival [4].Computer-Aided-Detection (CAD) systems for automatic pulmonary nodule detection were developed in the last two decades.The CAD system development is to fasten the interpretation of the CT images faster and more accurately, improving the screening program's cost-effectiveness [5].Without the gained ability for deep learning (DL) to learn features from data, those developments could not occur compared to the features made by hand relying on specialized knowledge in a particular domain.In present times, DL is increasingly being recognized as a means of advancing performance in various medical applications, as shown in Fig. 2. Using medical images to diagnose lung diseases using one of the DL models.Those developments enhanced clinicians' efficiency in examining and detecting certain medical cases [6].Technology development has ushered in a new era with challenges and possibilities.Technology, especially Artificial Intelligence (AI) [8], is vital in almost industry.They do this with the help of intelligent algorithms and powerful computers.However, each business is very different in how much it depends on computers.Technology is now an essential part of healthcare and medicine.It is possible by automating tasks and processes people used to do by hand.It saves time and makes it easier to understand diseases that are hard to explain.It could also decrease the cost of healthcare as a whole.In the past few decades, radiologists and physicians did most of the work of figuring out what medical images meant.Nevertheless, because so many diseases and experts can get tired, researchers and doctors are starting to use computer-assisted interventions [9].Learning from data is frequently utilized in CAD and examining medical images.In the realm of AI (i.e. machine learning), the classification process typically involves a sequence of procedures, including but not limited to feature extraction, learning, pre-processing, highly accurate feature selection, and classification.These steps are considered fundamental in traditional ML methodologies.The main thing used to make a choice is the chosen feature.However, it is possible that these things can be wrong and cannot be relied on to tell classes apart.Machines.DL helps the researcher to get features and classify them all at once.Fig. 1.Most commonly used datasets in previous articles for lung diseases Fig. 2. CT images detect diagnosing lung diseases using the convolutional neural networks (CNN) model [7] With the advent of Machine Learning (ML) and DL techniques [10], DL utilizing Artificial Neural Networks has emerged as one of the most vital feature extraction and diagnosis techniques.Despite multiple ML techniques for analyzing medical images in various domains, DL has emerged as the superior approach to looking at and understanding medical problems because it is more accurate [11].DL is a subset of ML that utilizes Deep Neural Networks (DNNs) with multiple inputs, outputs, and hidden layers.DL has gained widespread recognition for its 58% 39% 2% 1%

Covid-19
Lung cancer TB Pneumonia applications in various domains.Still, it is essential to analyze and interpret medical images to get around the limitations of image processing and ML methods, which is apparent in Fig. 3.However, these uses of DL are not very good at classifying things because there is not enough training data, and the models are not built well.DL models need data to work well [12].Note that the inadequate training data makes DL model network training inefficient.Before medical imaging uses CAD, these issues must be resolved.
The remaining sections of the paper are structured as follows: Section 2 discusses the types of lung disease images utilized in this research, while Section 3 presents the specific type of lung disease images examined in this study.Section 4 explains the process of image feature extraction using image processing and DL technique models, with a brief explanation of each model.In Section 5, previous works in this area are presented and compared.Section 6 introduces the mathematical models used to evaluate performance metrics for diagnosing lung diseases.Finally, Section 7 highlights the advantages of DL algorithms in diagnosing lung diseases.In Section 8, the challenges and limitations associated with the diagnosis of lung diseases are presented.Finally, the paper concludes in Section 9.

Type of Lung Disease
This section will briefly discuss the four diseases affecting the respiratory system and the lung, such as COVID-19, TB, Pneumonia and lung cancer.

COVID-19
Since December 2019, a new coronavirus illness has emerged, COVID-19 or 2019-nCoV.COVID-19 causes severe respiratory symptoms and widespread damage to the respiratory system.Coronavirus transmission can be reduced like any other illness, and patient survival rates can be improved via prompt identification and treatment.The clinician's go-to method for determining whether or not COVID-19 is present is the RT-PCR.According to the medical community, this method's positive rate is modest in the preliminary stages of this condition.As a result, physicians have to turn to another method to aid in rapidly identifying COVID-2019.As a result, the doctor looks more closely at the patient's CT scans and chest X-rays.Both techniques demonstrate that COVID-19 causes a unique lung alteration compared to other pneumonic diseases.Different AI algorithms and combinations of these algorithms are used to categorize these images.The results are contrasted and compared to ascertain the most effective approach, such as convolutional neural networks (CNN), k-nearest neighbor (k-NN), and Random Forest.

TB
TB has been a persistent problem among all infectious diseases, resulting in considerable morbidity and mortality.It makes TB more challenging to treat than other infectious diseases.Following a consistent decline in the number of new TB cases during the twentieth century, owing to improved social and environmental conditions, early detection, and the development of medication for treatment, stagnation and even a surge in new cases were observed in the mid-1980s.Improved social and environmental conditions, early detection, and the development of medication specifically for TB [13].Were the key factors contributing to the decrease in the number of new TB cases.The epidemiological shift can be attributed to a combination of factors, including an overall rise in developing nations, an increase in the number of people living with HIV and other forms of immunodeficiency, and an increase in the number of elderly patients whose immune systems have changed.Incomplete or insufficient treatment, the growth of multidrug-resistant TB, and a delayed diagnosis are other reasons that may be to fault.A delayed diagnosis is more common in people over the age of 65.The interaction between the organism and the host response determines the disease's progression and associated clinical and radiological patterns.Both of these aspects are dependent on each other.In the past, pulmonary TB was traditionally divided into two categories: primary TB, which affected children and post-primary TB, which affected adult patients.Primary TB was more common.However, adults, rather than children, were becoming infected with TB for the first time in developed nations [14].

Pneumonia
Pneumonia is a common lung illness in which a person's alveoli fill up with fluid and create a cloudy-like shape.It happens when a person has Pneumonia.There are two different kinds of Pneumonia, bacterial and viral, yet the X-ray patterns of each look highly similar [15].That requires guidance from computer-aided diagnostic tools.DL technology was utilized to diagnose Pneumonia in this investigation.Training the models, such as the VGG16 and Xception network, to better detect pneumonia cases.Thus, every network has unique capabilities to detect these diseases [16].

Lung cancer
Diagnostic imaging has shown lung nodules.Before CT scanning, chest X-ray lesions typically required resection.These tiny tumours (less than 3 cm) can cause symptoms such as sneezing, post-obstructive pneumonia, and hemoptysis.A lack of symptoms and solitary nodules in the lungs may also be present if they originate in the lung parenchyma.Lung nodules can be caused by several conditions besides malignant tumours [17].Doctors' experience diagnosing lung cancer may overlook some individuals and cause issues.Medical imaging diagnosis uses DL.This article designs CNN, DNN, and sparse auto-encoder deep neural networks for lung cancer calcification.Those networks are modified for benign and malignant lung nodules and used for CT image categorization [18].

Lung Diseases Image Type
Two distinct types of images are employed in training the model: CT-Scans and X-rays.These images are described in detail in Section 2. It is important to remember that additional imaging modalities, such as MRI scans, positron emission tomography, histopathology images, and sputum smear microscopy images, are all available.PET and MRI scans can detect disease and monitor treatment results.Unfortunately, only CT scans and X-rays are used in this review paper.

Lung diseases-based chest X-rays
X-rays are a standard diagnostic tool used by doctors to understand better and treat various health issues [19].A chest X-ray is the most common type of medical X-ray examination, and the resulting pictures may be used to examine organs like the heart, lungs, and airways, as well as the vertebrae and ribs in the patient's chest.Images captured by medical X-rays were previously subjected to photographic films, which must be developed before they can be seen.Digital X-rays are utilized to get around this issue [20].Fig. 4 shows medical images of patients with various lung diseases based on X-rays.

Lung diseases-based CT Scans
Radiography in a CT scan employs a series of pictures collected from different angles to construct cross-sectional pictures at different depths [25].The patient's tissues, organs, skeleton, and any anomalies can be shown in 2D on a flat screen or 3D by stacking the picture slices [18].When compared to X-rays, CT scan pictures provide more information.

Feature-Extraction
In order to perform a target task, it is commonly done to utilize a previously trained CNN model that has been trained on a large dataset.To accomplish this, it is necessary to preserve all model layers and adapt those fully connected to correspond with the new task.Convolutional layers are used to extract the feature.The information is then sent to a classifier built from fully connected layers and may be swapped out for a different job, or one of the ML classifiers, such as k-NN, support vector machine (SVM), etc., can be utilized.Instead of retraining the whole model, just the new classifier is trained [30].The key benefit of this strategy is the time savings gained by not having to re-run the model during each training iteration but rather only once on the new data.However, updating the training data on all is impossible with this method.Fig. 6 shows that convolution, pooling, and fully connected layers are essential elements of the CNN model.Sections 4.1 through 4.4 summarize the pre-trained models employed in this analysis.

MobileNetv2
Since its inception, MobileNet has undergone two major revisions, V1 and V2.The newer MobilenetV2 is a vast improvement over its predecessor [38].The MobileNet design is based on a typical convolution split into a depth-wise convolution and a 1x1 convolution, known as a pointwise convolution.Each input channel undergoes a single filter application in the depth-wise convolution.An 11 convolution is applied in the pointwise convolution to integrate the results of the depth-wise convolution.
In contrast to the one-step process of a conventional convolution, which involves mixing the inputs and filters to generate a new set of outputs, the depth-wise version of the algorithm separates these processes into separate layers.Reduced model size and calculation time are two significant benefits of this approach.Fig. 9 presents two DL models; MobileNet and MobileNetV2.

Previous Work
Based on clinical pictures, many ML and DL-based models have been built to identify lung disorders (CT scans and X-rays of the chest).When creating the methodology for this study, it is crucial to consider the most current research conducted in this area, which is discussed in the following review.

X-ray-based DL techniques
Rajaraman et al. [41] investigated an enhancing technique to improve the detection performance of TB lung diseases with a chest X-ray dataset that was divided into four sections (S, M, K, and I) by using CNN based on classifier models AlexNet, VGG16, GoogleNET and ResNet50.The findings were convincing and surpassed the existing cutting-edge.Ayan and Ünver [16]  Capizzi et al. [42] designed a special detection technique for lung cancer by using a probabilistic neural network (PNN) with fuzzy logic.This study used X-ray images of lung nodules to verify our model by computing the local variance on each pixel.Moreover, the suggested approach identifies and localizes possibly hazardous lung nodules.Finally, the test result has produced excellent outcomes compared to other models.Hou and Gwak [43] designed an automatic screening competitive technique for lung abnormalities images using a classified DL technique named knowledge distillation.The study used a dataset from the public ChestX-ray14 of thorax diseases.This strategy was designed to condense information from heavyweight teacher models such as ResNet-152 and DenseNet-121into lightweight student models such as MobileNet, VGG19, ResNet-59, and ResNet-50 or to self-train these student models in order to obtain poorly supervised multi-label lung disease diagnoses.
Tobias et al. [44] created a low-cost screening tool to classify normal or Pneumonia using a DL technique called CNN.The dataset used was chest X-rays of 6,555 images divided into 1,340 images of normal lung and 5215 images of Pneumonia.It presented them with a pre-trained model named MobileNetV2 that produced excellent results.Gite et al. [45] developed a new advanced segmentation technology to decrease the possibility of data leakage in diagnosing TB diseases by using DL instead of the classification architecture to focus on the vital region of the chest X-ray lung images, using U-Net + + and comparing them with neural network architecture such as SegNet, U-Net, and FCN.The generated result was tested on Shenzhen and Montgomery databases that showed U-Net + + achieved 98% accuracy; this indicates that it exceeds other architecture.Bhosale et al. [46] developed advanced IoT to reduce mortality and detect covid-19 and other obstructive lung diseases using a lightweight DL single model CNN technique using different chest X-ray images.The results showed encouraging output and enabled raspberry pi to detect lung disease.Table 1 compares the abovementioned research using X-ray-based DL techniques for lung disease recognition.Hasan et al. [53] designed algorithms to distinguish chest CT scans between three classes (COVID-19, Pneumonia, and healthy) using a combination of DL and handmade Q-deformed entropy (Q-DE) algorithms for extracting features.The pictures are pre-processed for the purpose of decreasing intensity differences across CT slices, and the CT lung scan was segmented using the histogram threshold.A recurrent neural network (RNN) based on long-short-term memory (LSTM), along with CNN and the Q-DE method, was used to classify the features extracted from each CT scan.In addition, an examination of variance ANOVA was employed to identify the pertinent characteristics.Luqy et al. [63] created three independent feature extraction techniques grey level co-occurrence matrix (GLCM), local binary pattern (LBP), and histogram-oriented gradient (HOG)were suggested to be combined.The authors applied principal component analysis on the output features generated by these techniques to decrease the number of features.The SVM is a machine-learning classifier used for classification.The data used to design the system (available on the online platform Kaggle) consists of 1100 CT scans and 1100 X-ray scans.The picture had 550 students in each class (with COVID-19 and no infected with COVID-19).When utilized with X-ray and CT scans pictures, the method provided an accuracy of 97% and 99%, respectively, for the binary categorization of COVID-19 and non-COVID-19.Table 2 compares the earlier research using CT scan-based DL techniques for lung disease identification.100.

Sound-based DL techniques
Fraiwan et al. [65] designed and evaluated a technique that notes respiratory and lung diseases from lung sounds with the raw acoustic signal dataset using deep CNN and bidirectional LSTM units called (CNN+BDLSTM).The K fold cross-validation scheme trained the model.The advanced model achieved high levels of quality, such as sensitivity and specificity of 98.43% and 99.69%, which opens up new possibilities for using DL applications in clinical programs.Pham et al. [66] presented a robust deep-learning technique for classifying anomalies and detecting illnesses in respiratory cycles that depend on sound using an auscultation record.With a dataset used of ICBHI, the model was trained using a CNN-MOE architecture.According to the results shown in the Table 3.
Aykanat et al. [67] This research represented a significant advance in the use of the medical stethoscope by developing it into an electronic one.The electronic stethoscope, in turn, contains some programs that help transmit the sounds of the respiratory system to the computer, where the computer then analyzes and diagnoses these sounds using the previously stored sounds.When the network is trained with the help of CNNbased DL by utilizing the characteristics of the Mel-frequency cepstral coefficients (MFCC) mixed with those of the SVM, some good results were discovered are listed in the table 3. Hazra et al. [68] This study focuses on developing methods for using sounds produced by the respiratory system in the early detection of diseases (such as bronchiectasis, Pneumonia and chronic bronchiolitis).this is done after extracting the acoustic characteristics, and it is fed into a file for acoustic data, where it is categorized using 2D CNN.Owing to the lack of layers required in training the algorithm, this model obtained results with high accuracy.
Kumar et al. [69] The suggested system can categorize early patient COVID-19 disease based on gathered chest X-ray images and Coswara cough (sound) samples from possibly sick individuals.Cough samples recorded are subjected to a series of pre-processing steps that use speech signal processing methods.The MFCC characteristics are retrieved with the assistance of deep CNNs.

Performance Matrices
In this section, we will count the most prevalent measures in previous studies, such as sensitivity, specificity, accuracy, F1-score, precision, ROC, and AUC, which are all important metrics to measure in order to obtain the best model for lung diseases detections [71].These performance metrics are commonly used to assess the impact of DL algorithms in diagnosing lung disorders.It is important to note that these metrics are not independent and can be influenced by various factors, including the size and quality of the dataset, the type of DL model used, and the specific application of the model.Therefore, it is essential to consider these metrics in conjunction with other factors.
A Confusion Matrix is a table with a specific shape that tabulates the algorithm's performance and efficiency based on its outcomes, with the actual values represented as rows in the matrix and the acceptable values displayed as columns and vice versa.The diagonally arranged values show the cases that are guessed correctly, while the others represent the wrong expectation, and the increase in the (diagonal) matrix values is better than the number of the lower values.Table 4 shows the Confusion Matrices cases.Accuracy can be found by dividing all collected samples' true positive and negative cases.
Sensitivity is determined by separating into equal parts the positive case forecasted correctly, the positive case appropriately forecasted, and the negative case forecasted improperly.This feature is useful.This statistic is helpful when an FN is further of a problem than an FP.
Specificity represents the capacity of the framework to detect negative instances accurately, and once all negative samples have been discovered, the model is good since it provides no unexpected outcomes.
Precision is calculated by dividing the number of true positive cases by the total number of positive cases that were predicted, including both correctly and falsely predicted positive cases.When an FP is more concerning than an FN, it is a helpful metric.
F1 score, while attempting to improve the model's accuracy, the sensitivity decreases.This metric may be viewed as a corresponding mean of accuracy and sensitivity, providing a comprehensive portrayal among both measures, with an F1 score reaching its highest rate at one and its poorest value at zero.
The receiver operator characteristic curve (ROC) is a rating metric for the binary classification task.At various threshold values, it shows the True-Positive Rate versus the False-Positive Rate [73].
The likelihood ratio (ROC) curve measures how effectively the equation separates two variables.Two-dimensional AUC measures the ROC curve area.This reflects model performance higher AUC.

Advantages of DL algorithm for lung diseases diagnosis
CNNs, DNNs, and GANs are DL models with advantages in diagnosing lung diseases based on medical imaging techniques such as X-rays and CT.The choice of an algorithm may depend on the specific application and the available data, and different algorithms may be more appropriate for different types of lung diseases or different imaging modalities.Here are some advantages of each type of network: CNNs advantages ▪ CNNs are specifically designed for image analysis tasks and are therefore well-suited for processing medical imaging data [74].▪ They can learn and identify complex features in medical images, making them helpful in identifying patterns and characteristics associated with different lung diseases [75].

▪
CNNs can handle large amounts of data and can be trained with relatively small amounts of labelled data [76].

DNNs advantages
▪ DNNs are more general-purpose models that can be applied to various medical imaging tasks, including lung disease diagnosis [77].

▪
They can learn complex nonlinear relationships between input data and output labels, making them well-suited for identifying subtle patterns and features in medical images [78].▪ DNNs can build more complex models than CNNs, improving lung disease diagnosis accuracy [79].

▪
GANs can be used to generate synthetic medical images, which can be used to augment existing datasets or to generate images with specific features or characteristics [80].

▪
GANs can generate medical images that simulate different lung diseases or imaging modalities, such as X-rays and CT [81].

▪
GANs can be used in conjunction with CNNs or DNNs to enhance the accuracy of lung disease identification by providing additional training data or generating more informative features [82].

Challenges and limitations for lung diseases diagnosis
DL has shown great promise in medical image analysis and has demonstrated high accuracy in diagnosing various diseases, including lung diseases.However, lung disease diagnosis based on DL still has several challenges and limitations.Some of these challenges and limitations are: ▪ Insufficient data: One of the significant challenges is the availability of labelled data, especially for rare or complex diseases.These networks require large amounts of high-quality data to learn from; without enough data, their performance can be limited [83].

▪
Overfitting: Another challenge is overfitting when the model grasps to distinguish distinct features to the training dataset but is not generalizable to new data.This can result in poor performance on new or unseen data [84].
▪ Interpretability: While DL models have shown excellent performance in many applications, the interpretability of these models can be challenging.It can be difficult to understand how the model arrived at its decision, which is vital for gaining the trust and acceptance of medical professionals [85].▪ Variability in medical images: Medical images can be highly variable, and the quality of the images can vary based on factors such as the imaging modality, the imaging device, and the patient's position.This variability can make it difficult for deep-learning models to diagnose lung diseases accurately [86].▪ Data imbalance: Imbalanced datasets, where significantly more examples of one class than another, can challenge DL models.This can result in biased models towards the majority class and may not perform well on the minority class [87].▪ Generalization: DL models sometimes have difficulty generalizing to new or unseen data.This is especially problematic in the medical field, where new and rare diseases may not have enough labelled data to train a model with high accuracy [88].
Overcoming these challenges will require continued research and development of new techniques and models that can improve the accuracy and interpretability of these models.

Conclusion
Analysis of medical images is a complex and growing topic that needs many studies.2, and Table 3.It was noted that choosing CNN models is the most common in diagnosing medical images.X-ray, CT, and sound.This review recommends using Fetcher extraction with an image processing technique to generate images from the original image to train the network.In addition, this review mentioned the advantages and disadvantages of each type of DL model (CNN, DNN, GANs).Finally, this review explains lung disease detection's current challenges and limitations.Overall, DL and image processing offer exciting possibilities for improving the detection and treatment of lung diseases.Ongoing research in this field will continue to refine these techniques and explore new applications for this powerful technology.

-
examined a new diagnostic technique for one of the most severe human lung infections, Pneumonia, using two commonly used CNN models.The network can provide clear X-ray images that radiologists can use as a successful diagnosis of the disease.The research used a data sample of 5856 frontal chest X-ray pictures and introduced them to two computerized support systems named VGG16 and Xception, which showed remarkable results in image classification.However, transfer learning and fine-tuning techniques were employed to train the two CNNs.The test outcomes have demonstrated that the VGG16 neural network far exceeds the Xception model in accuracy by 87%, pneumonia precision by 91%, specificity by 91%, and pneumonia F1 score by 90%.At the same time, the Xception network transcends the VGG16 network with regular precision of 86%, a sensitivity of 85%, and a pneumonia recall of 94%.The test result has shown that the VGG16 network is more accurate in the data than the Xception, but the Xception provided more successful results in detecting pneumonia cases.

Table 1 .
[52]arison of research using X-ray-based DL techniques for lung disease detectionWang et al.[48]designed a segmentation and early diagnosis technique for lung cancer using CAD and two CNN models.The study used a big-driven dataset of CT images and exported them to two models named AlexNet and ResNet that were used to train and test the data experiment.The training results displayed that AlexNet had high experimented than ResNet with an accuracy of 76% of AlexNet and 58% of ResNet.The advantage of this technique is that it can be used in pulmonary nodule extraction and automated recognition, which may save numerous lives.Still, it can also help to relieve medical resources, doctors, and patients.Yadav et al.[49]proposed an unsupervised DL system to diagnose lung disorders with chest X-ray and CT scans classification, such as Pneumonia and covid-19, by using a stacking classifier model that includes The model was trained by using a multi-layer GAN architecture called Lung-generative adversarial networks (GANs) the result achieved an excellent accuracy, precision, recall, F1-score.Tang et al.[50]developed a study of the segmentation of lung lob through the implementation of trained neurons, specifically DCNN.The study presented a dataset of 50 axial, sagittal, and coronal CT scans to LUNA16 and Tranche.The DL model was trained with dice loss, and the focal loss model and the convex hull model were trained with hybrid loss.Hu et al.[51]presented a weak supervised DL system CNN for rapid and automated diagnosis with data analysis.One of the diseases that might lead to death is COVID-19 infection using CT images from Community-acquired Pneumonia (CAP) and non-pneumonia clients.This system worked with a model that trained segmentation such as cross-validated and tested with actual manual truth TCIA dataset.Ardakani et al.[52] utilized several DL models to distinguish between patients with and without COVID-19.These networks were AlexNet, ResNet-18, VGG-19, GoogleNet, ResNet-101SqueezeNet, MobileNet-V2, ResNet-50, VGG-16 and Xception.This research comprised 108 individuals with COVID-19-positive lab results.In addition, 86 cases with atypical pneumonia were included.Their work adopted transfer learning to optimize CNNs for the datasets.The CNNs' input layer was a 60 x 60 x 1 infection patch, where 1020 image patches, including 510 COVID-19 and 510 Non-COVID-19, were extracted from CT slices by a radiologist specializing in infectious diseases.The findings indicated that ResNet-101 and Xception performed the best job among the ten networks.ResNet-101 could differentiate between COVID-19 and non-COVID-19 patients.
[58]etter classify multi-scale input pictures (128x128, 256x256, and 512x512), the MSCNN architecture learned feature representations.EfficientNetB0, a pre-trained network, was the backbone of three CNNs with the same structure but varied input scales for feature extraction.A dense, fully connected classifier with a sigmoid output activation function classified Hospitals provided 416 COVID-19 CT scans and 412 CAP CT scans.The system has 99.5% slice sensitivity and 95.6% slice specificity.Net, providing a 3D lung mask to a 3D DNN that predicted COVID-19 infection.313COVID-19and229non-COVID-19patientswerestudied.DeCoVNet was a three-stage model with two 3D residual blocks (ResBlocks), vanilla 3D convolution to extract image features, and a progressive classifier (three 3D convolutional layers, a fully connected layer, and a softmax activation function).The algorithm has 90.1% classification accuracy.Toqa et al.[59]developed a novel CNN for image classification, explicitly targeting the CT scans of COVID-19 patients.The model was trained and evaluated using 349 patients with COVID-19 photos and 195 without COVID-19 images.Augmentation of data was also employed to deal with the limited amount of the data samples.The suggested model used the softmax activation function at the output layer and had four convolutional blocks.Validation accuracy was 84.4%, while testing accuracy was 90.09% for this model.Compared to other pre-trained CNNs, the suggested model's performance confirmed its superiority in detecting capabilities (ResNet-50 and AlexNet).Adil et al.[47]developed the Deep Sense technique, a cross between a recurrent gated unit GRU neural network and a CNN.The model used three independent convolutional subnets to extract picture features, a merged convolutional subnet to process that output, and a GRU and output layer stacked in an 8-layer deep network to do classification.This model was tested on three distinct CT image benchmark datasets with varying validation and test splits, and it was compared to two other models built using an ANN.Compared to competing models using alternative data, the suggested model was shown to perform better in testing.Mucahid et al.[60]produced an ML COVID-19 classification scheme.It was divided into four subsets of differing sizes before being fed into the SVM classifier.To classify the four datasets, the author collected features by hand and fed them into (SVM), evaluating five different textural feature extraction methods.Several methods, including the Grey Level Co-Occurrence Matrix (GLCM), the Grey Level Run Length Matrix (GLRLM), the Grey-Level Size Zone Matrix (GLSZM), the Discrete Wavelet Transform (DWT), and the Local Directional Pattern Algorithm (LDP), were implemented.150CTimageswereused, and squares varying in size from 16 by 16 to 48 and 48 to 64 by 64 were selected randomly.The results showed that the GLSZM algorithm could achieve 98.71% accuracy using 5-fold cross-validation and a batch size of 100(32 x 32).Shuai et al.[58]DL algorithm to filter CT for COVID-19.A binary solution categorization is required for addressing both COVID-19 and typical pneumonia.The model's input photos were pre-processed to determine the ROI.The ROI was drawn from CT images using COVID-19 and pneumonia characteristics.Transfer learning was used to extract features from images using a modified Google-Net Inception v3 (M-Inception) network.This model extracted features that reduce ROI feature vector size.Categorization and prediction employ a fully connected network.One hundred eighty patients had typical viral Pneumonia, whereas 979 had COVID-19 (positive RT-PCR) and were treated at three facilities.The model was 89.5% accurate when internally validated but only 79.3% when externally validated.
[62]data samples comprised 118 CT scans images of COVID-19 patients, 107 healthy people, and 96 pneumonia patients.The suggested paradigm achieved a 99.68% classification rate.Pathak et al.[54]used Transfer Learning (TL) on the pre-trained network to create a COVID-19 classification model using 10-fold crossvalidation to avoid overfitting.A ResNet-50 network extracted chest CT image characteristics in the primary model design.Transfer learning was used to modify the deep layer starting parameters for the classification model.In addition, a CNN model consisting of a Softmax layer and four convolutional layers for activation was employed for classification.413COVID-19and439areregularorpneumonia-infectedCTscanswereutilized.The training accuracy of data was found to be 96.2%, while the accuracy of testing data was 93%.Yan et al.[55]created a classification system employing a multi-scale convolutional neural network (MSCNN) and tested its impact on slice and scan diagnosis.Li et al.[57]created a 3D DL technique, COVNet, to identify COVID-19 using 2D and 3D characteristics.U-Net neural network pre-processed the 3D CT scan to delineate the ROI.The COVNet, consisting of RestNet50, which forms the framework, collected features from a series of CT slices and performed a max-pooling operation to merge these features.A fully linked layer classifies in-depth features using the softmax activation function.COVID-19, CAP, and non-pneumonia were identified.The dataset contained 1292 COVID-19, 1735 CAP, and 1325 nonpneumonia CT images.Additionally, the authors employed Gradient-Weighted-Class Activation Mapping to visualize important regions identified by the model without any human annotation.The model detected COVID-19 with 90% sensitivity and 96% specificity.Wang et al.[58]used a lightly supervised DL system to identify COVID-19 from CT scan pictures without lesion annotation.The lung area was partitioned using a pre-trained U-Shah et al.[61]CT scans are automatically diagnosed and classed as normal or abnormal.As a result, clinicians rely on this approach to provide accurate and timely diagnoses, and these pictures are classified using DL models like the CTnet-10 model.It is due to the simplicity and speed with that it can be trained.However, the accuracy results are low compared to other models utilized, especially the VGG-19, where the results are 82.1% for the CTnet-10 and 94.52% for the VGG-19.Panwar et al.[62]used a transfer learning method to binary classify chest X-ray and CT scan images.The suggested technique used VGG-19 CNN pre-trained weights to assess in-depth features and incorporated five layers as a deep neural network classifier (a flat layer, two dense layers, average pooling, and a dropout layer).GradCAM used colour visualization to help radiologists understand the model.Chest X-ray and CT-scan images of four cases were included in the datasets: COVID-19, Non-COVID-19, Pneumonia, and Normal.CT scans included 1252 COVID-19 patients and 1230 non-COVID patients with other pulmonary disorders, while Xrays included 285 COVID-19 patients and 5856 pneumonia and regular patients.The model detected 95.61 percent of COVID-19 instances.

Table 2 .
Comparison of previous research employing CT scan-based DL techniques for lung disease identification

Table 3 .
Comparison of research using ultrasound-based DL techniques for lung disease detection

Table 4
One of the most critical problems facing researchers DL research is the accurate diagnosis of medical images.Therefore, the paper a review to present the medical image analysis techniques of respiratory diseases, representing pneumonia, TB, lung cancer, and Covid-19.According to the statistics provided by the World Health Organization, it was found that Covid-19 and Pneumonia have the highest rates among other diseases.DL and image processing have great potential for diagnosing and treating lung diseases.By leveraging large amounts of medical data, DL algorithms can analyze lung images and detect patterns not visible to the naked eye.It can lead to earlier detection of lung diseases, more accurate diagnosis, and personalized treatment plans.To classify the four diseases, the researchers used different DL models, most notably VGGNet, ResNet, and Xception, significantly in research to overcome this problem.Good results were shown, as mentioned in Table 1, Table