CNN-AE: Convolution Neural Network combined with Autoencoder approach to detect survival chance of COVID-19 patients

Abstract


Introduction
Nowadays medical centers are bloated with huge amount of data which are collected from patients.Medical biomarkers, demographic data and images modalities can help and support medical specialists to diagnose infectious diseases [1], Alzhemier [2], Parkinson [3], Coronary artery disease [4], etc.However, these data must be processed and analyzed in order to become usable information for the specialists.Automated solutions based on artificial intelligence have the potential to carry out the required process efficiently [5].
Recently, a new type of coronavirus (COVID- 19) has emerged which has caused many death worldwide [6][7][8][9].The virus outbreak has been observed for the first time in late 2019 [10,11].COVID-19 primarily aims the patients' lungs [12,13].Therefore, if the virus is not properly diagnosed in the early stages of infection, it can severely damage the lungs [14].Although the mortality rate of the virus is low, it must not be overlooked due to being highly contagious.The virus threat becomes more serious when the medical centers' resources cannot provide service to the large number of people who get infected each day [15].
Survival chance prediction of infected people is as important as early detection of the virus.Under resource scarcity, the medical centers can take into account patients' conditions and use the available resources wisely.Existing literature on COVID-19 detection has proven that deep neural networks are very effective for early detection of COVID-19 [16].Therefore, deep networks may be useful for survival chance prediction as well.In this study, we rely on clinical dataset including gender, age, blood type, etc. to perform diagnostic analysis on COVID-19 virus.To the best of our knowledge, this is the first time that a survival chance predictor is proposed for COVID-19 patients using clinical features.To evaluate the effectiveness of the proposed method, we compare its performance against a standard CNN trained on image data.Our contributions are as follows:  Survival chance prediction for COVID-19 patients based on clinical features  Preparing clinical dataset to predict survival chance of COVID-19 patients for the first time  Presenting careful analysis of the dataset characteristics including impact of features on mortality rate and correlation between each feature pair  Making our dataset publicly available  Combining Autoencoder with CNN to increase prediction accuracy  Proposing a data augmentation procedure to balance number of samples of different classes of dataset.Our data augmentation method is generic and applicable to any other dataset.
The remaining sections of the paper are organized as follows: related literature is reviewed in Section 2. Section 3 describes our dataset.Required background is briefly reviewed in Section 4. The proposed methodology is explained in Section 5. Experimental results are presented in section 6. Discussion, conclusion, and future works are presented in sections 7 and 8, respectively.This paper tries to predict survival chance of COVID-19 patients using clinical features.Therefore, we review COVID-19 detection methods which rely on clinical features.We also review methods on mortality estimation of infected patients.
To contain the COVID-19 threat as soon as possible, the researchers have tackled this virus from multiple directions.Some researchers have focused on fast and accurate detection of infected patients from noninfected ones.As an example, Wu et al. [17] extracted 11 vital blood indices through Random Forest (RF) method to design an assistant discrimination tool.Their method demonstrated an accuracy of 96.97% and 97.95% for the test set and cross-validation set respectively.The assistant tool was well-equipped to perform preliminary investigation of the suspected patients and suggested them quarantine and timely treatment.In another attempt, Rahman et al. [18] reviewed various studies on treatment, complications, seasonality, symptoms, clinical features and epidemiology of COVID-19 infection to assist the general people and medical staff in facilitating necessary guidance of the pandemic.Using a CNN they tried to detect infected patients in order to isolate them from the healthy ones.
In addition to distinguishing infected patients from non-infected ones, it is also important to determine whether infected ones have severe condition or not.Hence, Chen et al. [19] studied 148 severe and 214 non-severe COVID-19 patients from Wuhan, China using their laboratory test results and symptoms as features to devise a RF.The task of the RF was to classify COVID-19 patients to sever and non-sever types using the features.Using laboratory results and symptom as input, their model yielded accuracy over 90%.Some of the key features they identified were LDH, IL-6, absolute neutrophil count, D-Dimer, diabetes, gender, cardiovascular disease, hypertension and age.Some other researchers focused on mortality risk prediction of the patients.Gao et al. [20] proposed a mortality risk prediction model for COVID-19 (MRPMC) that applied clinical data to stratify patients by mortality risk that predicted mortality 20 days in advance.Their ensemble framework was based on four machine learning techniques including Neural Network (NN), Gradient Boosted Decision Tree [21], Support Vector Machine (SVM) and Logistic Regression.Their model demonstrated accurate and expeditious mortality risk stratification of COVID-19 patients.
Zhu et al. [22] presented a risk stratification score system and a multilayer perceptron (MLP) with six dense layers to predict mortality.78 clinical variables were identified and prediction performance was compared with pneumonia severity index (PSI), CURB-65 score, and COVID-19 severity score.They derived the top five predictors of mortality-lactate dehydrogenase, C-reactive protein, neutrophil:lymphocyte ratio, O 2 Index and D-dimer.Their model was proved to be effective in resourceconstrained and time-sensitive environment.
The power of XGBoost algorithm has also been leveraged for mortality risk prediction.As an example, Yan et al. [23] collected blood samples of 485 infected patients from China to detect key predictive biomarkers of COVID-19 mortality.They employed a XGBoost classifier which was able to predict the mortality of patients with 90% accuracy more than 10 days in advance.In another study, Bertsimas et al. [24] developed a data-driven mortality risk calculator for in-hospital patients.Laboratory, clinical, and demographic variables were accumulated at the time of hospital admission.Again, they applied XGBoost for mortality prediction on patients.Taking a different approach, Abdulaal et al. [25] devised a point-ofadmission mortality risk scoring system utilizing a MLP for COVID-19 patients.The network exploited patient specific features including present symptoms, smoking history, comorbidities and demographics and predicted the mortality risk based on them.The mortality prediction model demonstrated a specificity of 85.94%, sensitivity of 87.50% and accuracy of 86.25%.
Since the symptoms of different viruses may be similar to some extent, there has been an attempt to distinguish multiple viruses from each other in [26].To this end multiple classical machine learning algorithms were trained to classify textual clinical reports into four classes of SARS, ARDS, COVID-19, and Both (SARS, .Feature engineering has also been carried out using report length, Bag of words (BOW) and Term frequency/inverse document frequency (TF/IDF) techniques.Multinomial Naïve Bayes and Logistic regression outperformed other classifiers with a testing accuracy of 96.2%.The summary of the reviewed works are presented in Table 1.
Most of the existing literatures on COVID-19 rely on CT and X-ray images to achieve their objectives.Based on the review presented above, it is apparent that existing works based on clinical data are rather scarce.Hence we were motivated to present yet another study using clinical data for mortality risk assessment.The difference between our method and existing literature on mortality risk assessment is twofold.First, we present a new approach to carry out the assessment.Second, some of the clinical features that we have considered have never been used before which is why we release our dataset publicly.As will be discussed in the reset of the paper, clinical data are cost effective compared to CT images and classifiers trained on clinical data provide good performance which is almost as good as training on CT images.To justify this claim, we compare our method performance (trained on clinical data) with a standard CNN which is trained on CT images.[20] An Ensemble of NN, grad boosted decision tree, SVM, and logistic regression mortality risk prediction Zhu et al. [22] MLP mortality risk prediction Yan et al. [23] XGBoost classifier mortality risk prediction Bertsimas et al. [24] XGBoost classifier mortality risk prediction Abdulaal et al. [25] MLP mortality risk prediction Wu et al. [17] RF COVID-19 detection Rahman et al. [18] CNN COVID-19 detection Khanday et al. [26] Multinomial

Background
The proposed method consists of two modules which are classifier and data augmenter.The classification is carried out using a CNN.The data augmentation is realized using 10 autoencoders.In this section, we briefly review the main concepts of CNN and autoencoder.

CNN
CNNs are massively used in image based learning applications.Thanks to automatic feature extraction mechanism of CNNs, they can discover valuable information from training samples.CNNs are usually designed with several convolutional, pooling, and fully-connected layers [27].As illustrated in Figure 1, the feature extraction is done via convolving the input with convolutional kernels.The pooling layer reduces the computational volume of the network without making a noticeable change in the resolution of feature map.In CNNs, usually the size of the pooling layers decreases as the number of layers increases.Two of the most popular types of pooling layers are max pooling and average pooling [28].
Figure 1.A CNN schematic

Autoencoders
Autoencoders belong to the realm of unsupervised learning since they do not need labeled data for their training.In a nutshell, an autoencoder compresses input data to a lower dimensional latent space and then reconstructs the data by decompressing the latent space representation.Similar to Principle Component Analysis (PCA), autoencoders perform dimensionality reduction in the compression phase.However, unlike PCA that relies on linear transformation, autoencoders carry out nonlinear transformation using deep neural networks [29].The architecture of a typical autoencoder is presented in Figure 2.

Information gain
We review Information gain (IG) since in section 4, it is used to determine the degree to which each feature of our dataset contributes to the patients' death.IG calculates the entropy reduction due to splitting a dataset  based on a given value  of random variable : where () and (| = ) are entropy on dataset  and conditional entropy on  given  = , respectively.The conditional entropy is computed as: where  = ⊂  is the set of samples with variable  = .Moreover, | = | and || denote the cardinality of subset  = and set , respectively.In equation ( 1), the sum is computed over all possible values of .

Description of our clinical dataset
The dataset we have collected in this paper contains 320 cases, 300 of which are recovered patients and the remaining 20 cases represent deceased ones.The percentage of female cases is 55%.The mean age of the collected cases is 49.5 years old and standard deviation is 18.5.The patients have referred to Tehran Omid hospital in Iran from 2020-03-03 to 2020-04-21.Ethical approval of these data was also obtained from the Omid hospital.Tobacco {Yes, No} Considering that our dataset has not been released before, it is vital to assess the degree to which each dataset feature contributes to patients' death.Such an analysis provides researchers with valuable insight on the characteristics of the collected data.Various feature selection methods are available to determine the weight of each feature in classification of dataset samples.We chose information gain [31] which is one of the widely used feature selection methods [32].In Figure 3, the importance of each feature (i.e.information gain) is shown as a bar.The age has a much larger information gain (0.149) compared to other features.Thus age is not included in Figure 3 to make the comparison of other features importance easier.According to the bar chart, (after age) cancer, heart, and kidney diseases are the second, third, and forth most important feature related to patients' death.Hence, it is clear that the patients with poor health conditions are more vulnerable against COVID-19.It is worth noting that Figure 3 does not include the features with information gain zero.
We have also inspected the interplay between the dataset features to find out the potential correlation between them.To this end the grid in Figure 4 is presented which can be thought as a heat map showing positive/negative correlation between features.Each cell (, ) in the grid of Figure 4 represents the correlation of features in i-th row and j-th column.As the cell color approaches red the positive correlation between the feature pairs is higher.For example Anosmia and Ageusia have high positive correlation which means they are usually observed simultaneously.Anosmia is the loss of ability to smell and Ageusia is the inability to taste with tongue.

Proposed Methodology
In this study, the survival chance prediction of COVID-19 patients who have referred to Omid hospital in Tehran is investigated.The classification is based on features obtained from patients' information.In the database provided, the number of recovered people is 300 and the number of deceased ones is 20.It is clear that the number of recovered patients is much higher than the number of deceased ones.In order to classify accurately, it is necessary to balance recovered to deceased ratio of dataset samples.To do so, the number of instances of the lower class is increased such that the number of data in both classes is approximately equal.In order to increase the number of data of deceased people, an autoencoder model has been used.To carry out the data augmentation, the 20 samples of deceased class are fed to the autoencoder to undergo compression and decompression routine.The output of this process is 20 reconstructed samples which are similar to the original ones but not identical.Therefore, we have augmented the original 20 samples with 20 reconstructed ones.Training the autoencoder 10 times using different training and validation sets yields 10 autoencoders with similar architecture but different parameters.Each of the 10 autoencoders can generate 20 reconstructed deceased samples yielding overall reconstructed samples of 200 which will be added to the original deceased samples.To gain a better insight on what the autoencoders do, a sample vector before and after reconstruction is presented in table Table 3.For the majority of "1" components of input vector , the autoencoder has outputted values near 1 as components/elements of the reconstructed vector .Similarly most of the reconstructed components corresponding to original components "0" have values near "0" which shows that the reconstruction process is sound.The details of the augmentation process are explained in more detail in sub-section 5.1.It is worth noting that our augmentation procedure is generic and can be applied to any other dataset.

Implementation details of CNN-AE
The proposed method (CNN-AE) consists of multiple steps which are summarized in Figure 5.The detailed explanation of these steps is presented below: 1. Design 10 autoencoders with identical configuration but different initial parameters for data augmentation.To implement the proposed method, we used Python language and Keras library which has TensorFlow backend.In this study, the dataset contains 320 samples of infected cases.The number of recovered cases is 300 and the remaining 20 cases represent the deceased ones.Moreover, we have also generated 200 reconstructed deceased cases to balance the recovered to deceased ratio of our dataset.After the reconstruction phase, our dataset contained 520 cases.We have used 10-fold cross validation.Moreover, 80% of 9 of the folds are used for training and the remaining 20% are used for validation.The implementation details of CNN and AE are illustrated in Figure 6 and Figure 7, respectively.

Experiments
In this section, the experimental results are presented.The implementation details of CNN and autoencoders are explained in section 6.1.We report the performance of the proposed method (CNN-AE) and compare it with a CNN in section 6.2.

Experiments setup/details
Our experiments consist of two scenarios.In the first scenario, our method (CNN-AE) is compared with a standard CNN which is trained on clinical data.The architecture of the CNN is presented in Table 4.To have a fair comparison, we use the same CNN architecture in our method.Moreover, implementation details of autoencoders used in CNN-AE are presented in Table 5.

Experimental results
In this section, we seek to answer two important questions about the proposed method.First, we compare our method performance with a standard CNN which is trained on clinical data.This experiment reveals the effect of the proposed data augmentation technique using multiple autoencoders.Next, we train a standard CNN for the same purpose (survival chance prediction) but using CT images.This experiment determines how well CT images can represent patients' survival chance using a CNN as the predictor.

Inspecting data augmentation approach
As mentioned in section 5.1, we use 10 autoencoders to augment the available dataset.Data augmentation is critical to successful training when the number of samples from different classes is unbalanced.Data imbalance can bring any powerful classifier to its knees even a state of the art CNN which is the motivation for our data augmentation technique.
To investigate the effectiveness of our data augmentation procedure, we have trained a CNN on original dataset and our CNN-AE on the augmented one.The original dataset has only 20 samples with deceased label while the recovered samples are 300.Comparing 300 with 20 reveals severe data imbalance from which CNN has suffered during training as shown in Table 7.On the other hand, using augmented dataset with 300 recovered samples and 220 deceased ones has facilitated the CNN training leading to better accuracy ( Table 7).Moreover, AUC measure of CNN-AE is almost twice the AUC of CNN.The specificity measure of CNN is mostly zero which stems from the fact that CNN has been unable to distinguish deceased samples from the recovered ones due to insufficient number of deceased samples in the original dataset.Looking at  7, CNN-AE method with average accuracy of 96.05% outperforms CNN method with average accuracy of 92.49%.Moreover, thanks to the augmented data, our method has been able to reduce training/validation loss faster than CNN which is evident in Figure 9. Similarly, CNN-AE has managed to reach higher accuracy faster than CNN which can be seen in plots of Figure 10.During training, our method exhibits much variation in validation plots compared to CNN.The reason is that CNN has quickly overfit to small number of deceased samples but CNN-AE had to deal with more versatile augmented samples.Therefore, CNN-AE had harder time during training but could gain better overall performance.

Comparison with image-based CNN
In this section we evaluate the performance of a standard CNN which is trained on a dataset of CT images.The CT images have been taken from the same patients for whom the clinical dataset has been collected.Therefore, the results of this section reveal how well a CNN trained on CT images performs compared to a CNN trained on clinical data.It is worth noting that most of the conducted experiments in the COVID-19 literature revolve around classifying infected and non-infected people using CT images.However, this section sheds some light on how well a CNN can predict survival chance of already infected patients based on CT images.
The dataset consists of 2822 CT image of recovered patients and 2269 CT images of deceased ones.The CT image dataset size is much greater than clinical dataset size.The reason is that the CT dataset contains multiple images per each patient.Since the number of samples of the two classes in the dataset is almost balanced, we did not apply our data augmentation technique to the CT dataset.Moreover, having multiple images per each patient acts as some form of data augmentation.This is not the case for clinical dataset where for each patient only one value for each feature is present.

Comparison with other methods
We have also compared our method performance with other commonly used classification methods such as Naïve Bayes classifier, random forest, logistic regression, and support vector machine (SVM).To get a better understanding of the working mechanism of these methods, first we review them briefly and next present the experimental results for them.

Naïve Bayes
Naïve Bayes classifier is based on Bayesian network which is a probabilistic graphical model.Given the observed event, a Bayesian network predicts the likelihood of the cause of that event.For example, given symptoms, the network can predict probabilities of the various diseases.

Radom forest
As the name implies, random forest is a set of decision trees which are used in an ensemble setup to get highly accurate classification results and avoid overfitting at the same time.Each of the trees is built using a random subset of available training data.The subset samples are draw randomly with replacement from the training data.Moreover tree nodes are split based on a random subset of the training data features.
During the testing phase, decision trees of the forest provide their judgment on the given test sample hich are averaged to yield the final prediction [33,34].

Methods performance
In this section we report experimental result using other classification methods such as random forest and Naïve Bayes.We also investigate the effect of using the proposed data augmentation technique during the training.Such an experiment demonstrates that our data augmentation method is generic and applicable with other methods as well.The performance statistics are presented with 95% confidence interval in Table 9.The confidence intervals are computed based on 10-fold cross validation.For each method we have performed two experiments.First, each of the methods is trained on the original dataset (without augmentation) the results of which are presented in first and third rows of Table 9.Once again the methods are trained but on the augmented dataset.The augmentation is carried out using our autoencoder based approach.The results are shown in second and forth rows of Table 9.Both of random forest and Naïve Bayes have clearly benefitted from the augmentation performed on the training dataset.The augmented dataset has led to significant improvement of Naïve Bayes accuracy.The augmentation has mildly improved random forest accuracy but yielded much higher AUC.The last row of Table 9 is identical to the last row of Table 7 which is restated here for ease of reference.It is evident that the proposed method CNN-AE has outperformed other methods both in terms of accuracy and AUC.

Discussion
This paper focuses on survival chance prediction for COVID-19 patients.We performed experiments using clinical dataset as well as CT images dataset.Recall that the CT image dataset size is almost ten times of clinical dataset size.Yet, the CNN trained on clinical data could almost perform as well as the CNN trained on CT data which encourages using clinical data as an alternative for CT images.
Another aspect that may encourage using clinical training samples is the data collection cost.Preparing CT data may require high-end facilities which lead to increased data collection cost.Moreover, the required facilities to prepare CT data may not be available in deprived areas whereas the requirements to measure clinical data such as blood pressure, fever, CRP, etc. are accessible.
The proposed method can detect the severity of patients' condition based on clinical data and take preventive actions to minimize the mortality rate.As reviewed in section 2, few methods have studied mortality rate prediction using clinical data.Moreover, the existing methods have used features different than the ones we have used in our experiments.Therefore, the proposed method sheds some light on unexplored aspects of COVID-19 virus.In addition to the proposed method, our dataset can be considered as the second contribution of this paper since it is a good resource for further medical research.To the best of our knowledge we are the first to analyze the importance of the dataset features and their correlation as presented in Figure 3 and Figure 4. Using our dataset, the experts can study the relation between patients' medical conditions (e.g.blood pressure, diabetes) and COVID-19 death likelihood.As a result, the medical experts can exercise more caution during treatment of patients who are more likely to die due to their medical condition.As the information gain values in Figure 3 suggest, there is a strong relation between mortality rate of COVID-19 patients and having critical diseases such as cancer, kidney and heart diseases.On the other hand, mild symptoms and/or diseases like dyspnea, conjunctivitis, and asthma are less likely to contribute to the mortality rate.
Like any other classification approach, the proposed method has some shortcomings as well.Due to using multiple autoencoders in the data augmentation phase, the training time of our method is longer than a standard CNN.Moreover, standard CNNs receive a single image sample as input and perform feature extraction automatically.On the contrary, we have collected multiple clinical features by hand for each patient which is harder to manage.Some of the features in our dataset have been asked directly from the patient so it is possible that the patient provides wrong information.

Conclusions and Future works
In this paper we investigated the possibility of using clinical data to train a CNN for survival chance prediction of COVID-19 patients.The experiments have been conducted on a clinical dataset which we have collected.Our dataset consists of new features that have not been studied before.We have made the dataset publicly available and analyzed its features using information gain and correlation.Such an analysis can aid the potential researchers and practitioners with their work on COVID-19 virus.To reduce the data imbalance of our dataset, we proposed a novel data augmentation method based on autoencoders.
Our data augmentation approach is generic and applicable to other datasets as well.The effect of using data augmentation was investigated to evaluate its effectiveness.Using augmented data for training, CNN-AE has been able to reach accuracy 96.05±1.48%,recall 98.00 ± 1.33%, and specificity 93.13 ± 2.52%.However, training a CNN on dataset without augmentation has yielded accuracy 92.49±2.75%,recall 95.4±0.88%, and specificity 96.9±3.73%.It is clear that CNN-AE has benefitted the data augmentation and outperformed CNN.
We repeated the CNN training but this time using CT images of the same patients from clinical dataset.
Comparing the performance of training on clinical data with training on CT data reveals that clinical data can be used as an alternative for CT images.
As future work, it is important to collect more data to be able to assess the proposed approach better.It is also important to investigate the use of other data augmentation methods and compare the results with our data augmentation.
Features effect on mortality rate based on information gain Correlation between dataset features The steps for implementing the proposed method.
Implementation details of CNN: circles with "B" letter represent batch normalization layers and circles with "D" letter represent dropout layers with probability 0.5.

Figure 2 .
Figure 2. Autoencoder architecture: high dimension input data is encoded (compressed) to form latent (hidden) space which has a lower dimension compared to the original input.The latent representation is reconstructed (decoded) to yield decompressed output.

Figure 3 .
Figure 3. Features effect on mortality rate based on information gain

2 .
Train each of the 10 autoencoders on 300 samples representing the recovered patients.We want to have 10 models with different parameters at the end of the training.To this end, we divide the 300 samples to ten groups of 30 samples {  ,  = 1,2, … ,10} where   is the j-th group of samples.To train i-th model,   is set aside for validation and the nine remaining groups {  ,  ∈ {1,2, … ,10} − {}} (270 samples) are used for training.Recall that each model is initialized with different parameters, trained on partially different training samples, and validated on totally different validation set.Therefore, the proposed training procedure will yield 10 different autoencoders.3. The 20 deceased samples are fed to each of the 10 trained autoencoders.The samples undergo the compression and decompression routine of autoencoders.The decompression procedure is lossy so the 20 reconstructed samples (after decompression) are not identical to the original ones.Moreover, the 10 trained autoencoders exhibit different behavior on the same input data since their parameters are different from each other.Therefore, feeding the same 20 samples to the 10 autoencoders will yield 200 new samples which belong to the deceased class.The motivation behind the explained procedure is data augmentation to remedy lack of enough samples for the deceased class.4. The 200 reconstructed samples are attached to 320 original ones to yield a dataset of 520 samples. 5. Design a CNN model to classify 520 samples as recovered or deceased.6. Train the CNN model using all 520 samples.We apply 10-fold cross validation during the training.Hence, the training sample size is 468 (samples of 9 folds) and the test sample size is 52 (samples of 1 fold).7. Use the trained CNN to classify the test data.

Figure 5 .
Figure 5.The steps for implementing the proposed method.

Figure 6 .
Figure 6.Implementation details of CNN: circles with "B" letter represent batch normalization layers and circles with "D" letter represent dropout layers with probability 0.5.

Figure 9 .
Figure 9. Loss plots of CNN and CNN-AE methods during training on our dataset

Figures Figure 1 A CNN schematic Figure 2
Figures

Figure 7 Implementation details of Autoencoder 8 Implementation
Figure 7

Figure 9 Loss
Figure 9

Table 1 .
Summary of reviewed literature

Table 2 .
To gather the cases, patients' history collected by doctors, questionnaire filled by patients, laboratory test, vital sign measurement, and imaging investigation were used.The description of the dataset features is presented in Table2.Our dataset is publicly available in[30].Institutional approval was acquired for the use of patient datasets used in research studies for diagnostic and therapeutic purposes.Approval was granted on the grounds of existing datasets.Informed consent was obtained from all participants of this study.All methods were carried out in accordance with relevant guidelines and regulations.Description of dataset features which are used for classification

Table 4 .
Implementation details of the CNN which has been trained on clinical data.

Table 5 .
Autoencoder implementation details phase of our experiments, we compare CNN-AE (trained on clinical data) with a standard CNN which is trained on image data.The CNN architecture is presented in Figure8.After multiple trials, we ended up with the best set of CNN hyperparameters which are presented in Table6.

Table 6 .
Implementation details of the CNN which has been trained on image data Figure 8. Implementation details of the CNN which is trained on CT images.

Table 7 ,
it is clear that CNN-AE training takes more time which is due to the time it takes to train 10 autoencoders required for data augmentation.

Table 7 .
Comparison of CNN and CNN-AE using different evaluation metrics based on 10-fold cross validation.

Table 8 .
Results of CNN trained on CT images

Table 9 .
Performance metrics for RF and Naïve Bayes with and without autoencoder based data augmentation