Brain Tumor Classification Using Hybrid Single Image Super-Resolution Technique With ResNext101_32× 8d and VGG19 Pre-Trained Models

High-quality images acquired from medical devices can be utilized to aid diagnosis and detection of various diseases. However, such images can be very expensive to acquire and difficult to store, and the process of diagnosis can consume significant time. Automatic diagnosis based on artificial intelligence (AI) techniques can contribute significantly to overcoming the cost and time issues. Pre-trained deep learning models can present an effective solution to medical image classification. In this paper, we propose two such models, ResNext101_ $32\times 8\text{d}$ and VGG19 to classify two types of brain tumor: pituitary and glioma The proposed models are applied to a dataset consisting of 1,800 MRI images comprising in two classes of diagnoses; glioma tumor and pituitary tumor. A single-image super-resolution (SISR) technique is applied to the MRI images to classify and enhance their basic features, enabling the proposed models to enhance particular aspects of the MRI images and assist the training process of the models. These models are implemented using PyTorch and TensorFlow frameworks with hyper-parameter tuning, and data augmentation. Experimentally, receiver operating characteristic curve (ROCC), the error matrix, Precision, and Recall are used to analyze the performance of the proposed model. Results obtained demonstrate that VGG19 and ResNext101_ $32\times 8\text{d}$ achieved testing accuracies of 99.98% and 100%, and loss rates of 0.0120 and 0.108, respectively. The F1-score, Precision, Recall, and the area under the ROC for VGG19 were 99.89%, 99.90%, 99.89%, and 100%, respectively, while for the ResNext101_ $32\times 8\text{d}$ they were all 100%. The proposed models when applied to MRI images to provide a quick and accurate approach to distinguishing between patients with pituitary and glioma tumors, and could aid doctors and radiologists in the screening of patients with brain tumors.


I. INTRODUCTION
An important issue in medicine is classification of brain tumors [1], [2] to decide on treatment type.This is a challenging issue because tumor cells are heterogeneous by nature.Doctors, of course, are important in the diagnosis of this disease, but would benefit from a helpful tool to aid rapid diagnosis [3].Today, diagnostic systems based on computer-aided technology provide an effective means for diagnosing brain tumors via magnetic resonance imaging (MRI) [4].MRI is the most common technique used for classifying brain tumors, because of its high image quality [5].Artificial intelligence (AI) is also being considered as a key enabler to assist in resolving issues around brain tumor classification.In particular, the development of high-performance deep learning models (DLMs) with high levels of accuracy would be a significant step towards a fast, high-precision method for detection and diagnosis of brain tumors in patients.
Machine learning and deep learning models are being utilized to classify and diagnose brain tumors [6], but, the low accuracy of existing models is a challenge that needs to be addressed [7].Deep convolutional neural networks (CNNs) are playing an effective role in detecting the presence of brain tumors and diagnosing tumor type via medical imaging using MRI, which has been used to successfully detect patients infected with brain tumors.However, a disadvantage of these networks is the long time required to train the models [8].Low accuracy can be significantly improved by using pre-trained DLMs, such as VGG [9], DenseNet [10], ResNet [11], GoogLeNet [12], AlexNet [13], MobileNet [14], and EfficientNet [15] which can also reduce training time.Speech recognition [16] language modeling [17], human activity recognition [18], and image processing [19] are already making good use of such models.Pre-trained models are more convenient since they require less training time.However, achieving a sufficiently high accuracy in classifying brain tumors remains a significant challenge.
This paper reports on implementing ResNext101_32 × 8d and VGG19 models to enhance the accuracy of detection and classification of brain tumor using hyper-parameter optimization.These two models are based on a transfer learning process, and have the advantage of architectural simplicity, which can reduce computational cost (training time).The models were tested and trained on a dataset available in the Kaggle repository.In addition, a data augmentation technique was employed on the dataset used in order to overcome its deficiency of the available medical images.
The primary contributions of this paper are: • Implementation of robust ResNext101_32 × 8d and VGG19 to distinguish between glioma and pituitary brain images/cases for fast and more accurate diagnoses; • Using ResNext101_32 × 8d and VGG19 with hyper-parameter optimization to successfully achieve substantially greater test accuracy for a graphics processing unit (GPU); • Applying a single image super-resolution (SISR) technique for the two models based on a generative adversarial network (GAN) algorithm to improve the resolution of the MRI images; • Evaluating the performance of the models with alternative evaluation measures based on an assessment of the quality of images in a dataset of 1,800 MRI images of the brain; • Improving VGG19 accuracy by using k-fold crossvalidation; • Comparing the results achieved by the proposed models for the MRI image classification with different state-of-the-art models.
The remainder of this paper is as follows: Section II presents a brief review of similar work.Section III introduces the theoretical background to transfer learning.Section IV presents the proposed DLMs, dataset description, and performance metrics utilized.Section V presents the experimental results.
The results are discussed in Section VI.Section VII offers conclusions and suggestions for further work.

II. RELATED WORK
The literature contains reports of a number of DL models for classifying brain tumors using MRI images of the brain [20], [21], [22], [23], but a satisfactory level of accuracy has not always been achieved.In [20] Researchers [24], [25], [26], [27], and [28] presented machine and deep learning architectures to detect brain tumors in patients.In [24], Cheng et al., introduced an SVM model and attained a 91.28% accuracy with a dataset of 1,426 glioma and 930 pituitary images.A model using CNN architecture was applied to a dataset of 3,064 MRI images [25] and attained 84.19% accuracy.Kaplan et al., [26] applied a k-nearest neighbor (KNN) model with nLBP feature extraction approach, and achieved 95.56% accuracy.Pashaei et al., [27] implemented a CNN model with an extreme learning machine (ELM) method and trained using 70% of a dataset that contained 3,064 brain tumor cases.The model was then used to assess the other 30% and attained an accuracy of 93.68%.In [28], Zacharaki et al., proposed SVM-KNN models for brain tumor classification, which VOLUME 11, 2023 were applied to a dataset of 102 MRI images, achieving an overall accuracy of 85%.
Kurup et al., [29] utilized the architecture of a CapsNet model based on three classes to predict the presence of brain tumors.It applied capsule neural networks to a set of data comprising 3,170 images with an accuracy of 92.6%.In [30], Das et al., applied a CNN for brain tumor detection with a dataset containing three classes of 1,426 glioma, 708 meningioma, and 930 pituitary cases.The accuracy and precision achieved were 94.39% and 93.33%, respectively.Ullah et al., [31] proposed to classify brain tumors using MRI images via an artificial neural network (ANN), and attained a testing accuracy of 95.80%.In [32], Huang et al., used a CNN to classify brain tumors and attained an overall accuracy of 95.49%.In [33], Kalaiselvi et al., diagnosed brain tumors using a CNN with 96.00% accuracy.Li et al., [34] used a hidden Markov model (HMM) for real-time classification of brain tumors and achieved an accuracy of 96.88%.Noreen et al., [35] introduced two models: Xception and Inception-V3, and used 3,064 images to achieve accuracies of 93.79% and 94.34%, respectively.
Rehman et al. [36] classified microscopic brain tumors using a 3D CNN.This model was applied to the BraTS datasets, using the k-fold cross-validation method, achieving a 96.67% accuracy.In [37], Rehman et al., implemented GoogLeNet, VGGNet, and AlexNet models, utilized 3,064 images, and classified three types of tumors.Sharif et al., [38] presented an InceptionV3 model for brain tumor recognition, with an achieved 93.7% accuracy.In [39], Muzammil et al., proposed a multimodal image fusion algorithm for diagnoses using MRI images.It was applied with a convolutional sparse coding method which used the entropy theorem to assess the performance of the algorithm.Maqsood et al., [40] investigated brain tumor detection using U-NET CNN and fuzzy logic algorithms, but did not quote a success rate.Their literature review showed that despite significant research into detecting the presence of glioma or pituitary tumors in the brain via MRI images, improvements more accurate methods are still needed.

III. BACKGROUND FOR TRANSFER LEARNING
Transfer learning is a machine learning technique whereby a model trained for one purpose is subsequently used for another.The weights from the pre-trained model are applied as a starting point for training using a different dataset for a different issue.A significant advantage of transfer learning is a faster training/learning process.Another advantage is achieving higher performance with less data due to the weights obtained from the previous training being used.A higher performance can be achieved by the addition of a fully connected layer to the existing model.
Transfer learning is usually associated with relatively small datasets, such as biomedical images.If a DLM is trained ab initio, the training process requires a large amount of data and time.Thus, it is convenient to utilize a pre-trained model and fine-tune its performance to accelerate the training process.There are many successful pre-trained CNN models used in the classification of medical images, including ResNext101_32 × 8d and VGG19 [41], which were pre-trained using the ImageNet dataset, which consists of 1.3 × 10 6 RGB images of 224 × 224 pixels and with 1000 classes.It is proposed to use VGG19 and ResNext101_32 × 8d as pre-trained DLMs for transfer learning to classify brain tumor types.

IV. METHODOLOGY
ResNext101_32 × 8d and VGG19 models are employed here to classify two categories of brain tumors and were applied to a dataset consisting of 1,800 MRI brain images.These models were chosen due to their robust performance and to be convenient for processing spatial data [42].

A. THE PROPOSED SISR TECHNIQUE WITH RESNEXT101_32 × 8dI
In this work, before classification by the ResNext101_32×8d model, a SISR technique was applied to the chosen MRI brain tumor images.Fig. 1 illustrates the architecture of the SISR technique with ResNext101_32 × 8d model.The SISR is based on a GAN algorithm to produce high-resolution images.The SISR consists of two phases: the first is the generator, and the second is the discriminator.The generator comprises an input layer with a shape of 64 × 64 × 3 and a kernel size of 3, followed by an up-sampling block containing a convolutional layer and a Parametric Rectified Linear Unit (PReLU) layer, this is followed by a residual block which repeats every sixteen iterations and comprises: a convolutional layer, a Batch-normalization layer, a PReLU layer, another convolutional layer, another Batch-normalization layer, and an Add layer.These are followed by three layers: convolutional, Batch-normalization, and Add, followed by two blocks Convolutional, Lambda, and PReLU.Finally, the output layer is activated by a sigmoid activation function.
Every convolutional layer (conv-layer_has 3 × 3 kernels and 64 filters.The discriminator model comprises an input layer, a convolutional layer, and a ReLU layer, followed by seven repeated blocks containing a convolutional layer, a Batch-normalization layer, and a ReLU, this is followed by a block that includes a flatten layer, a dense layer, and ReLU, this block is repeated three times.Next, there is a dense layer and an output layer with a sigmoid activation function.The dimensions of the high-resolution (HR) images are 256 × 256 × 3 and the low-resolution (LR) images are generated from the high-resolution images.The LR images have dimensions of 64 × 64 × 3. So, the HR is divided by a factor of 4 in order to obtain the dimensions of the LR images.For the SISR technique, only 1700 images were used from the three thousand MRI images available due to the generator model used in the training process being slow.The images were normalized to the range from 1 to −1. 1,550,659 trainable parameters were available for the generator model-based SISR technique.

B. THE PROPOSED SISR TECHNIQUE WITH VISUAL GEOMETRY (VGG)-19
In this work, a SISR technique was also used on the MRI brain tumor images before classification by VGG19.Fig. 2 displays the architecture of the proposed VGG19 model, which comprises 19 layers: three fully connected layers and sixteen 2-D conv-layers, each of which is followed by a 2-D MPL.Training VGG19 takes less time than other pre-trained models while also having high classification accuracy.
As previously, the input MRI images have equal width and height of 224 pixels.First, a 2-D convolutional layer was applied separately to each input image, with a ReLU activation function to extract spatial features.This layer has 64 filters, a kernel with 3 × 3 matrix shape, followed by another convolutional layer with 64 filters with a ReLU function.To make the convolution output less complex an MPL with a 2 × 2 matrix, carried out a downsampling procedure.
Third, there were two conv-layers having 128 filters, a kernel with a 3 × 3 matrix, and utilizing a ReLU function.These added layers enable the VGG19 to discern higher-level features that might have been missed in the previous convlayers.Fourth, an MPL with a 2 × 2 pool size is followed by four 2-D conv-layers having a configuration of 256 filters, a kernel with a 3 × 3 matrix, followed by an MPL and then four 2-D conv-layers having a configuration of 512 filters with an MPL.Fifth, four more 2-D conv-layers having a configuration of 1024 filters, followed by an MPL.Sixth, two fully connected layers are configured with 4096 neurons and a ReLU activation function, followed by a fully connected layer with 1000 neurons.Finally, the output is reduced to just two classes by the application of a softmax activation function.
The difference between predicted and true values for the VGG19 was obtained using a binary cross-entropy loss function, with the loss function calculated using Eq. ( 1), with N the number of classes, y the true output label, and ŷ the predicted label.Again an Adam algorithm was used as an optimizer [43].It was found that a batch size of 30 with 50 training epochs were best to train the VGG19 model for classification of brain tumors.The total number of trainable parameters for this model was 2,325,568.Table 1 presents a summary of the layers for the VGG19 model.C. DATASET DESCRIPTION Fig. 4 shows four of the MRI brain images from the dataset supplied by the Kaggle repository and used for classification [44].There were 1,800 MRI brain images of two classes, 900 glioma tumors and 900 pituitary tumors.Each image was resized into 224 × 224 pixels, and then normalized by rescaling the pixels from [0, 255] to [0, 1].To minimize over-fitting, for VGG19, three data augmentation techniques were used to increase the original The datasets were divided, the training sets were 75% for the VGG19 model and 85% for the ResNext101_32 × 8d model.Thus the corresponding test sets used to assess the two models were, respectively, 25% and 15% of the datasets.

D. EVALUATION METRICS
In this work, several well-known evaluation metrics were used to analyze model performance [45], [46]: Accuracy, Precision, Recall, F1-score, and the area under the receiver operating characteristic curve (ROCC).The ROCC is a means of comparing the accuracy of different classification models, to demonstrate the ability of a test to correctly identify those images with a tumor.The ROCC is a graph of the True Positive Rate (TPR -the images that were correctly diagnosed as having a tumor as a proportion of all images that did show a tumor) against the False Positive Rate (FPR -the images that were incorrectly diagnosed as having a tumor as a proportion of all images that did not show a tumor).The ranges for both TPR and FPR are between 0.0 and 0.1.
If TP = number of diseased samples correctly identified, TN = number of healthy samples correctly identified, FN = the number of samples that were diseased but falsely diagnosed as healthy, and FP = the number of samples that were healthy and falsely diagnosed as diseased.Using these metrics, the total number of images in the dataset is (TP + TN + FP + FN), and the total number correctly identified is (TP + TN).
Accuracy of the model is the ratio of images accurately identified to the total number of images, see Eq. ( 2).Precision is the ratio of the number of images correctly diagnosed in a particular class, e.g., TP, to the total number of images diagnosed as in that class (TP+FP), see Eq. (3).Recall or sensitivity, is the ratio of number of images correctly diagnosed (TP), to the total number of correctly identified MRI images of both classes (TP+FN), see Eq. ( 4).It is the probability of a positive test if the patient has a glioma tumor.
F1-score, or balanced F1-measure, is the harmonic mean of the Recall and Precision weighted by a factor of 2, see Eq. ( 5).The F1-score includes both FN and FP, so it can sometimes be a more useful metric than Accuracy.

V. EXPERIMENTAL RESULTS
Python in a Google Colab environment, with P100 GPU and 25 GB RAM memory was used to implement the proposed models.performance contained no errors.The total number of glioma and pituitary tumors which taken for the testing error matrix is 270.Table 2 presents the F1-score, Precision, and Recall for VGG19, by which to assess its performance for the dataset utilized.The values for F1-score, Precision, and Recall for the glioma and pituitary classes were, respectively, 99.92% and 99.86%, 99.07% and 99.74%, and 99.78% and 99.00%.The macro-average is determined by computing an evaluation metric independently for each class and then taking the mean.For the F1-score, Precision and Recall the respective macroaverages were: 99.89%, 99.90%, and 99.89%.We note, see Table 2, that the corresponding weighted averages had the same values.
Fig. 9 shows histograms for the glioma and pituitary images for VGG19.Precision-recall and ROCCs are presented in Figs. 10 and 11.Precision-recall curves present the precision rate as a function of the recall rate.Fig. 10 shows the precision-recall curves for VGG19 for both glioma and pituitary classes.The values of the areas under the precision-recall curves for both classes are 1.00 or 100%,   giving the corresponding values of the macro-average for both precision and recall of 1.00.Fig. 11 shows the ROCC for VGG19, and it can be seen that the values of the areas for both classes are 1.00 or 100%, meaning the macro-average ROCC area is also 1.00.These results imply that VGG19 doesn't cause errors.Table 3 presents performance data for the proposed VGG19 model using k-fold cross-validation, where k represents the number of equal partitions into which the data is divided.Here k = 5, one for validation and four for training.The model was trained five times using the different partitions, each time with an epoch number of 50.The average loss rate and accuracy rate for the model were 0.0113 and 99.35% respectively.Therefore, VGG19 achieved a high performance when using 5-cross-validation, avoiding bias in the results by using a suitable allocation of test and training datasets.Table 4 presents F1-score, Precision, and Recall metrics for ResNext101_32 × 8d, by which to assess its relative performance for the dataset used.For both the glioma and pituitary classes, the F1-score, Precision, and Recall were all 100%, as were the macro-averages and weighted averages.
Fig. 16 presents the precision-recall curves for ResNext101 _32 × 8d for both glioma and pituitary.The values of areas under the precision-recall curves for both classes are 1.00 or 100%.Therefore, the macro-average precisionrecall curve area is 1.00.Fig. 17 shows the ROCCs for ResNext101_32×8d, and, again, the values of the areas under

VI. DISCUSSION
The results presented above show that the ResNext101_32 × 8d and VGG19 models have a very high accuracy with low loss rate when trained and tested.Precision-recall curves, error matrices, and ROCCs demonstrated that the proposed models can accurately classify brain tumors.Table 5 is a comparison of the accuracy achieved by the VGG19 and ResNext101_32 × 8d models with previously published results.We see that the testing accuracies of the proposed models are noticeably higher than those achieved by the models listed in references [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34], and [35].The higher performance achieved is due to batch and kernel sizes, the fine-tuning of the models' hyper-parameters, loss activation functions, optimizer type, pool size, number of neurons utilized in the conv-layers, and number of training epochs.
For VGG19, setting the number of training epochs to 50, and batch size to 30, and using a softmax activation function with kernel and pool sizes of the convolutional and max-polling layers adjusted to 3 × 3 and 2 × 2 filters, respectively, VGG19 achieved a test accuracy of 99.89%.However, when the batch size was changed to 64, the training epochs reconfigured to 40, with a sigmoid activation function, and kernel and pool sizes of the convolutional and max-polling layers set to 5 × 5 and 3 × 3 filters, respectively, the VGG19 testing accuracy reached only 95.78%.We conclude that the parametric settings can significantly enhance the results.
A GridSearchCV method was used to automatically compute the optimum values of the hyper-parameters to ensure the models achieved optimal performance.

VII. CONCLUSION
This paper reports the application of ResNext101_32×8d and VGG19 DLMs to classify patients with glioma and pituitary tumors based on brain MRI images.The models were trained and assessed using a dataset of 900 each of glioma and pituitary images.In addition, a single image super-resolution (SISR) technique was applied to the MRI images to improve their resolution before classification using ResNext101_32 × 8d and VGG19.The SISR is based on a GAN algorithm and evaluated using MS-SSIM, PSNR, and SSIM metrics.The MS-SSIM was 96.39%, the PSNR was 29.30 dB, and the SSIM rate was 0.847.Experimental assessment of the accuracy of the VGG19 and ResNext101_32 × 8d models, showed the accuracy realized was 99.89% and 100% respectively, with respective test loss rates of 0.0120 and 0.0108.
The error matrix, F1-score, Precision, Recall, area under the precision-recall curve, and the ROCC have been presented and the models' performances evaluated.The VGG19 model's F1-score was 99.89%, its precision score was 99.90%, and the achieved recall was 99.89%.The corresponding precision-recall curves for the VGG19, for both glioma and pituitary tumors was 100%.The area under the ROCC is 100% for both classes for the VGG19.
The ResNext101_32 × 8d model's F1-score, precision, and recall were all 100%.The achieved areas under the ROC and precision-recall curves were 100% for both classes glioma tumor and pituitary tumor.Models such as these assist specialist doctors by providing a fast identification of patients with brain tumors, which makes these models useful tools for rapid screening and providing support for medical diagnoses.
The hyper-parameters of both ResNext101_32 × 8d and VGG19 i.e., batch and kernel sizes, fine-tuning the models' hyper-parameters, loss activation functions, optimizer type, pool size, number of neurons used in the conv-layers and number of training epochs.were found to substantially impact the accuracy of the results.Best performance depends on achieving the optimal settings for these parameters.
The obtained results demonstrate that the pre-trained ResNext101_32 × 8d and VGG19 models achieved high performance when classifying brain tumors.To assess the performance of VGG19 in terms of testing accuracy and loss 5-fold cross-validation was used.Both the ResNext101_32 × 8d and VGG19 models can be applied to MRI medical images to speed up diagnosis for the benefit of both patients and doctors.
Future work should include applying the models proposed here to more brain MRI images, possibly also adding other pre-trained DLMs, such as ResNet-18 and AlexNet, to the utilized dataset.

FIGURE 1 .
FIGURE 1.The workflow of the proposed SISR technique with ResNext101_32 × 8d model.

Fig. 3
shows the workflow for the proposed VGG19 model.There are six steps, the first was uploading the MRI images dataset used, which were separated into images for testing and training.The second step was pre-processing the MRI images i.e., image normalization.The third step was to define the number of training epochs.The fourth was the training of the model using the designated MRI images via a fitting function.The fifth step was to test the prediction capacity of the VGG19 model using the MRI test images.The final step was to evaluate the performance of the model using different metrics on the MRI test images.

FIGURE 3 .
FIGURE 3. The workflow of the VGG19.

FIGURE 4 .
FIGURE 4. Four samples of brain MRI images from the dataset used: (a) and (c) Glioma Tumors, (b) and (d) Pituitary Tumors.

FIGURE 5 .
FIGURE 5. Training and validation accuracy curves for VGG19.

Fig. 5
Fig. 5 presents the training/learning and validation accuracies obtained for VGG19.The blue line symbolizes the training accuracy, which increases with increase in the number of epochs and approaches 100% after 50 epochs.The brown curve shows the validation accuracy, commencing at 97.56% and rising to 99.89% after 50 epochs.The training stopped at 50 epochs because the learning curve started to overfit.The number of training epochs was tuned for the highest training/ validation accuracy.

FIGURE 6 .
FIGURE 6. Training and validation loss curves for VGG19.

Fig. 6
Fig. 6 shows the training/ validation loss curves for the VGG19 model, where the score of 0.0 would indicate perfect learning with no mistakes.Both losses continuously decreased as the number of epochs increased with the training loss reaching 0.0030 after 50 epochs, and validation loss commencing at 0.110 and declining to 0.0120.

Figs. 7
Figs. 7 and 8 present the error and normalized error matrices obtained from use of the VGG19.This matrix is used to evaluate model performance when classifying two classes, here using the MRI test dataset and comparing the predicted/true class outputs.The dark purple blocks on the matrix in Fig. 7 represent classification accuracy, while the values outside the blocks represent error values.Here, the error matrix shows, respectively, 129 and 141 true positives for the two classes of tumor, glioma, and pituitary.The normalized error matrix for VGG19 shows classification accuracy of 1.0 (100%) for glioma and pituitary classes, with zero classification error in both cases.The VGG19 model

FIGURE 9 .
FIGURE 9. Histograms of Glioma and Pituitary images.

Fig. 14
Fig. 14 shows the error matrix obtained using ResNext101_ 32 × 8d.The dark purple blocks show the classification accuracy.Here, the error matrix shows, respectively, 128 and 142 true positives for the two classes of tumor: glioma, and pituitary.The normalized error matrix for ResNext101_32 × 8d is illustrated in Fig. 15.It has a classification accuracy of 1.0 ''100%'' for glioma and pituitary classes, with zero classification errors.Table4presents F1-score, Precision, and Recall metrics for ResNext101_32 × 8d, by which to assess its relative performance for the dataset used.For both the glioma and pituitary classes, the F1-score, Precision, and Recall were all 100%, as were the macro-averages and weighted averages.Fig.16presents the precision-recall curves for ResNext101 _32 × 8d for both glioma and pituitary.The values of areas under the precision-recall curves for both classes are 1.00 or 100%.Therefore, the macro-average precisionrecall curve area is 1.00.Fig.17shows the ROCCs for ResNext101_32×8d, and, again, the values of the areas under

TABLE 4 .
Classification of precision, recall, and F1-score for the proposed ResNext101_32 × 8d model.

FIGURE 17 .TABLE 5 .
FIGURE 17. ROC curves for ResNext101_32 × 8d.TABLE 5. Comparison of the accuracy of proposed and previous models.

FIGURE 18 .
FIGURE 18. Results of high-resolution images: (a) High-Resolution MRI image, (b) Histogram of high-resolution image.

FIGURE 19 .
FIGURE 19.Results of low-resolution images: (a) Low-Resolution MRI image, (b) Histogram of low-resolution image.

FIGURE 20 .
FIGURE 20.Results of super-resolution images: (a) Super-Resolution MRI image, (b) Histogram of super-resolution image.

TABLE 1 .
The summary of the layers of the Vgg19 model.

TABLE 2 .
Classification of precision, recall, and F1-score for the Vgg19 model.

TABLE 3 .
Performance of the Vgg19 model for different k-fold cross-validations.