A Novel Approach to Classify Brain Tumor with an Effective Transfer Learning based Deep Learning Model

Abstract

World's deadliest disease is brain tumor. Misdiagnosed cancers and inadequate treatment reduce survival. However, magnetic resonance imaging (MRI) is utilized for tumor analysis, but the enormous number of pictures produced by MRI makes it time-consuming and difficult to diagnose a patient just because of its complex nature, putting their life at risk. Thus, accurately detecting early brain cancers manually is difficult. We need an autonomous, intelligent system to detect brain cancers early and accurately. This study proposes a pre-trained EfficientNetb4 model with an adjusttable learning rate and custom callback to efficiently classify tumors. The proposed methodology improves the quality and quantity of the publicly accessible Br35h dataset by applying data augmentation techniques. The suggested model had 99.67% accuracy on Br35h data and 0.33% miss categorization. This method, however, had 99.87% accuracy on supplemented data and 0.13% miss categorization. The system achieves 99.33%, 99.97%, 99.93%, 99.33%, 0.66%, and 99.66% for Br35h in sensitivity, specificity, precision, NPV, FOR, and F1-score. The suggested model had 99.97% sensitivity, 99.74% specificity, 99.74% accuracy, 100% NPV, 0.66% FOR, and 99.66% F1-score for the Br35h-augmented dataset. Further, the proposed model also achieved FNR and FPR of 0 and 0.26%, respectively, and the augmented Br35H dataset achieved FNR and FPR of 0.66% and 0%, respectively. The 5Flod cross-validation also found that the model achieved an overall validation accuracy of 98.606% with a loss of 0.224%. Regarding results, the proposed system shows superior performance to other state-of-the art approaches.

Keywords:
Deep learning; Transfer learning; Callbacks; Learning rate; Medical image analysis; Br35h; EfficientNetB4

HIGHLIGHTS

To reduce the computational time, we pre-processed the publicly available Br35h brain tumor dataset.

To overcome Overfitting data augmentation techniques applied to enhance data set usability.

Custom call-backs and an adjustable learning rate used for better tumor prediction automatically, adjust the learning rate with comparison of the threshold value by seeing validation lose.

A pre-trained customized EffiientNetB4 model is proposed for brain tumor detection by adding some hyperparameters and layers, as well as saving resources in terms of memory consumption.

INTRODUCTION

Two major components of the body are brain and spinal cord, that control all central nervous system of the body, hence, any harm in any of them cause lead to death of anybody [¹] and it considered extremely dangerous for the life of a person. So, a tumor in the brain also causes danger to life, and it’s an abnormal growth of cells and tissues in the human brain or spinal cord. Due to its severity, a tumor can be classified into two classes: malignant and benign [²]. Because tumors are varied, they have different symptoms and ways to be found. According to the WHO, benign tumors are non-cancerous and rarely reappear after recovery, while malignant tumors can spread from one part of the body to another and endanger human life [³].

A benign tumor [⁴] comes from brain cells, while a malignant tumor might arise from other regions of the body or the brain. Benign tumors grow slowly and stay in borders or identical parts of where they originate, while malignant tumors grow quickly and are more dangerous because they spread rot to adjacent tissues [⁵]. Because of its creeping nature, it's hard to identify early, yet earlier diagnosis can improve human life.

Early brain tumor discovery and diagnosis can enhance treatment and survival rates with other treatments. Despite this, doctors or pharmaceutical staff can diagnose the patient using various methods, such as physical brain examination, biopsy, digital screening, and others, but all of these can be done with medical imaging modalities like X-rays, CT scans, and MRIs. Brain X-rays [⁶] detect broken bones, while CT scans [⁷] show the brain's interior structures. MRI is one of the most advanced imaging techniques that classifies or separates tissues or cell sections [⁸], helping detect brain tumors as malignant. Because of MRI, pharmaceutical professionals can easily see soft tissue features.

So, manually detecting and diagnosing brain tumors led to missed predictions and diagnoses, which is dangerous for patients and time-consuming. To save patients' lives, an automatic and accurate system that can use a better approach for earlier and more accurate brain tumor detection is needed. Computing techniques are needed to make this possible. Medical personnel would struggle to find accurate, early, and possible patient diagnoses without technology. Brain tumor diagnosis by parametric personnel is widespread due to systemic healthcare issues.

Many computing methods have been researched for brain tumor detection. Machine learning, deep learning, and artificial intelligence were mostly used with medical imaging detail to classify tumors, such as artificial neural network model (ANN) [¹⁰], support vector machine (SVM) [¹¹], K-nearest neighbor (KNN) [¹²], decision Tree [¹³], etc. Machine learning-based techniques rely on hand-crafted feature extraction from images [¹⁴], which requires a lot of time and memory for big data sets [¹⁵]. In deep learning-based models like ResNet [¹⁶], AlexNet [¹⁷], VGG16 [¹⁸], MobileNet [¹⁹], features from medical imaging are automatically extracted and used to evaluate the final outcome for brain tumor classification.

Medical imaging still struggles to detect brain tumors early and improve ML (machine learning) and DL (deep learning) models. Thus, we suggested an automatic and accurate approach for early brain tumor detection utilizing MRI-based images by applying a deep learning-based modified model.

Further, the organization this research article is as follows: the introduction and background is discussed in Section 1. Section 2 demonstrate the reviewed techniques and their performance comparisons. The proposed methodology is explained in Section 3, and Section 4 shows the achieved results on the proposed approach. Conclusion and future work exists in Section 5.

LITERATURE REVIEW

Khan and coauthors [²⁰] proposed a solution to categorized brain tumor into cancerous and non-cancerous class by using transfer learning approach. The experiment was conducted by using VGG19 and MobileNetV2 pretrained model on publicly available dataset which contained 2513, and 2087 unhealthy and healthy images respectively. Proposed system achieved 97% and 91% accuracy on MobileNetV2, and VGG19 model respectively.

Tahir and coauthors, [²¹] presented a research based on Machine learning algorithms, adaptive back propagation neural network (ABPNN) and SVM along with fuzzy logic approach for the detection of tumor in brain. Experiment was performed using BraTs dataset, and revealed that ABPNN achieved 98.67% training accuracy with 96.72% testing accuracy, on the other hand training and testing accuracy was achieved by SVM is 98.48%, and 97.70% respectively. The overall accuracy of proposed model after applying fuzzy logic achieved 98.79%, and 97.81% training and testing accuracy respectively.

Saravanan and coauthors, [²²] proposed a neighboring network limitation technique along with a CNN database learning approach for the classification of brain tumor images in the medical domain. The proposed architecture consists of multilayered metadata learning with the integration of CNN layers for accurate transfer of information. Implementation was carried out on two different publicly available datasets, BraTs and REMBRANDT, and it was claimed that the proposed system works more efficiently than other existing techniques.

Siddiqi and coauthors [²³] presented another brain MRI image classification investigation using Harvard Medical School and OASIS datasets. The suggested method uses Z-values to extract the best features from MRI images and separate image classes. SVM classifiers are used to predict image label class after feature extraction and selection. They claimed their procedure was the most accurate and efficient compared to current methods.

Sangeetha and coauthors [²⁴] proposed a multi convolutional transfer learning base mechanism to detect tumor in brain on a small dataset. For this initial baseline was chosen for the extraction of features from 3D brain images, while tumor was classified using convolutional autoencoder mechanism. System was performed almost 1.5% better in achieving accuracy in term of detection of tumor in brain.

Hamza and coauthors [²⁵] presented a research work that was based on an efficient and optimal deep learning model for tumor classification in brain MRI images. Noise was removed from the dataset by applying a bilateral filter, and skull stripping was applied for the pre-processing phase. Morphological segmentation was carried out for detecting affected regions, while the Xception model with sooty tern optimization was applied for the extraction of features. Tumor images are further classified using attention-based long- and short-term memory.

Reza and coauthors [²⁶] designed a strategy to classify brain tumor in MRI images using a deep convolutional neural network. They proposed VGG16 architecture on a dataset that contained 10,153 MRI-based images with three different tumor classes, such as gliomas, meningiomas, and pituitary tumors. The proposed architecture achieved precisions of 99.4%, 96.7%, and 100% for the glioma, meningioma, and pituitary tumor classes, respectively, with an overall accuracy of 99.5%.

Another piece of research was done by Hossain and coauthors [²⁷], who provided a solution to classify the brain tumor into benign and malignant classes. To recreate microwave images of the brain into six different classes, they proposed an eight-layer lightweight classifier with a self-organized operational neural network. The dataset contains 1320 images that were chosen for experiments and reconstructed into 13200 images. The proposed model achieved almost 98% accuracy in terms of outcomes.

Vision transformers ensembling-based tumor categorization in Brain MRI images was proposed by Tummala and coauthors [²⁸]. This was done by applying pretrained and fine-tuned ImageNet ViT models with multiple versions to the publically available Figshare dataset of 3064 T1W images. The L/32 model variant scored 98.2% accuracy on 384 × 384 photos, while other Vit model variants achieved 98.7% accuracy on 224 × 224 images.

Samee and coauthors [²⁹] proposed a hybrid transfer learning model for the diagnosis of brain tumor-related medical imaging in three distinct classes (glioma, meningioma, and pituitary). They did so by combining five layers of GoogleNet and ten layers of AlexNet models that extract features and classify lesions from the publicly available CE-MRI dataset. The proposed model outperforms other models in terms of accuracy and sensitivity, with an accuracy of 99.51 percent and a sensitivity of 98.90 percent

Vision transformers ensembling-based tumor categorization in Brain MRI images was proposed by Tummala and coauthors [²⁸]. This was done by applying pretrained and fine-tuned ImageNet ViT models with multiple versions to the publically available Figshare dataset of 3064 T1W images. The L/32 model variant scored 98.2% accuracy on 384 × 384 photos, while other Vit model variants achieved 98.7% accuracy on 224 × 224 images.

Gomez-Guzman and coauthors [³⁰] presented research for the classification of brain tumors using CNN on MRI images. For this, they used seven CNN architectures, one of them was generic, and the other six models were pretrained. The Msoud dataset was used, which contained 7023 images of different tumor categories that were contained in different publicly available datasets. In comparison and based on our findings, the InceptionV3 model achieved better results by achieving 97.12% accuracy.

Usmani and coauthors [³¹] proposed a research model with the effect of interactive batch normalization and learning rate with the use of a transfer learning approach for classifying tumors in brain. The authors used ResNet18, ResNet50, and RestNet101 with different optimizers (adam, rmsprop, and sgd). As a result, their proposed model achieved 99.56% accuracy after applying effective hyperparameter tuning effects.

Uysal and coauthors [³²] proposed a solution for the multiclass classification of tumor in the brain using various deep learning approaches (ResNet, RegNet, and Vision Transformer). The two different publicly available datasets were chosen for experiments, pre-processing was done by applying contrast limited adaptive histogram equalization, and the dataset was randomly divided for training, testing, and validation up to (80, 10, and 10%), respectively. The results demonstrated that fine-tuned proposed models had accuracy levels greater than 90%.

Jibon and coauthors [³³] proposed an improved classification mechanism for the classification of brain MRI cancerous or non-cancerous images by using CNN with Log polar Transformation. Log polar transformation was used to extract features from images that were rotated and distorted, while CNN was applied for the classification of tumors in distorted images. The results showed that the proposed techniques achieved up to 96% accurate result in plain and distorted images.

Ozdemir and coauthors [⁴²] presented a research work based on the MTAP model for brain tumor classification, whose main purpose was to enhance the accuracy of the brain tumor detection system. For this, they used the Figshare dataset, and on that, they applied preprocessing steps to overcome the class imbalance issue, and their model achieved an accuracy of 99.69%.

Another research work was proposed by Ozdemir and coauthors [⁴³] for brain tumor classification on the Figshare dataset, on which they applied a customized CNN algorithm to classify tumors, and the dataset was generalized by a data augmentation approach for better results and achieved an accuracy of 98.69%.

Atasever and coauthors [⁴⁴], presented a comprehensive survey on medical image analysis by focusing on the on the transfer learning-based concept of deep learning models. For this, they considered 125 research articles from 2017 to 2021, grouping them into 8 regions, and discussing their limitations, challenges, opportunities, and future works.

Khaliki and coauthors [⁴⁵], presented a research work based on brain tumor detection on a publicly available dataset and applied four different pretrained models, including InceptionV3, EfficientNetB4, VGG16, VGG19, and a customized 3-layer CNN model. Out of all the models, the VGG-16 model produced better results and achieved an accuracy of 98% and f1-score of 97%.

Table 1. Shows the summary of above discussed research work, and summarize them in limitation, approaches that have been used on different datasets for the improvements in accuracy.

Table 1
Comparison of Brain Tumor Detection Literature Survey and their Performance.

Authors	Techniques	Dataset	Accuracy/evaluation	Remarks/Limitations
[20]	Pretrained VGG19	Kaggle (2513 cancerous, 2087 healthy)	91%	Need to improve accuracy
	Pretrained MobileNetV2		97%
[21]	Adaptive backpropagation neural network and SVM with fuzzy logic	Brats	ABPNN = 96.72%,	Need to improve accuracy / complex architecture.
			SVM = 97.70%
			Fuzzy Logic = 97.81%
[22]	CNN database learning along with neighboring network limitation	BRATS and REMBRANDT	N. A	Accuracy is not mentioned and Complex architecture
[23]	Forward, backward recursion model with SVM	Harvard medical school dataset	N. A	Need improvement in complex architecture and enhancement in accuracy required
[24]	Multi convolutional transfer learning with autoencoder	N. A	93±1%	Need to improve accuracy, and dataset is not mentioned
[25]	EADL-BTMIC, STO, Xception, and ALSTM	N. A	99.06%	Complex preprocessing and architecture, dataset is not mentioned
[26]	Modified VGG-16	10153MRI images	99.5%	Need to overcome overfitting issue
[27]	MBINet Model using (Self-ONN)	1320 RMB images	98%	Use better optimization algorithm and make sure to overcome complexity
[28]	Vision Transformer Ensembling	Figshare dataset 3064 MRI images	98.2%, 98.7%	Used different architecture variant for classification and difficult to implement on different datasets.
[29]	Hybrid transfer learning (GoogleNet and AlexNet)	Online publicly available dataset	99.51%	NA
[3]	Hybrid transfer learning (GoogleNet and CNN)	Br35h dataset	99.51%	Try to overcome overfitting issue
[30]	Generic CNN model with ither six pretrained model	Msoud (combinantion of Sartaj, Figshare, and Br35h dataset) 7023 images	97.12%	Need to improve accuracy
[31]	Cartesian product matrix-based approach with ResNet 18	3064 MRI images from publicly available	99.56%	Complex structure
[32]	fine-tuned ResNet, RegNet, and Vision Transformer (ViT)	NA	>90%	Need to improve accuracy
[33]	Convolutional Neural Network and Log-Polar Transformation	Formed by itself	96%	Need to improve dataset and accuracy

The above table shows the comparison of different reviewed approaches regarding their approach, dataset details, results, and limitations. Many approaches used complex architectures, achieving less accurate results, and some found overfitting issues. Still, there is a need for improvements in the results regarding using a well-suited dataset and a fine-tuned model that produces better results and performs accurate classification.

MATERIAL AND METHODS

This section shows the detailed proposed methodology, from data acquisition to final result evaluation. Major steps in methodology are data preprocessing, including data augmentation, data resizing and scaling, and class labeling, then further acquired and augmented datasets are passed to a pretrained deep learning model for the classification of brain MRI images in the Br35h dataset. System performance is measured in terms of accuracy, specificity, sensitivity, recall, F1-score, ROC, etc. The Figure 1 shows the workflow of the proposed model.

Figure 1
Pipelined for Classification of Brain Tumor.

A) Data Set Acquisition

The proposed methodology was applied to the Br35H dataset, which can be downloaded from Kaggle and used for the classification of brain tumors. A lot of different researchers and programmers used this dataset, and this dataset has a usability rating of 8.75 out of 10 on Kaggle [⁴⁶], which shows it’s a very authentic and reasonable dataset for brain tumor detection and classification. There was a total of 3000 images in the collection, with 1500 categorized as tumors and another 1500 categorized as healthy. All the pictures have been categorized and labelled. The Br35h dataset's training, testing, and validation procedures are detailed in Table 2, and the dataset's pictures, labelled with yes or no, are displayed in Figure 2.

Table 2
Br35H dataset description along with training, testing, and validation ratio.

Tumor Class	Total Images	Training Samples	Validation Samples	Testing Samples	Class Label
Healthy	1500	1200	150	150	Yes
Cancerous	1500	1200	150	150	no

Figure 2
Br35H MRI images of Yes and No Tumor Class

B) Data Set Preprocessing and Augmentation

After acquiring the images, the first step applied was to resize the image to 300 × 300. After resizing, labels were converted to NumPy arrays for better compatibility with models and efficient computation. After correct interpreting and handling of class information, one hot encoding is applied to labels to format them, and input pixels normalize between 0 and 1 for better training the model. Data augmentation is the process of generating more data from existing data. The main advantage of this technique is that it improves model prediction and accuracy while reducing data overfitting by generating more images from the dataset with different angles and shapes, which helps the model learn actual information from images on different aspects. For data augmentation, we augmented the images on a 90-degree rotation with a width and height shift range of 0.2, which is used to specify the upper bound of the fractional total width and height for images from 0.0 to 1.0. Further, we moved the dataset images both vertically and horizontally with a shear and zoom range of 0.2, and based on it, we have created more samples for better learning of features from the images and trained an augmented and simple dataset for evaluating the performance of the transfer learning-based proposed EficientNetB4 model. Further, Figure 3 shows the augmented images of the healthy and tumor classes, and Table 3 shows the description of the augmented Br35h dataset along with training, testing, and validation statistics, and Table 4 shows the details of the data augmentation hyperparameters.

Table 3
Augmented Br35H dataset description along training, testing, and validation ratio.

Tumor Class	Total Images	Training Samples	Validation Samples	Testing Samples	Class Label
Healthy	3900	3128	391	391	Healthy
Cancerous	3900	3128	391	391	Tumor

Table 4
Data Augmentation Hyperparameters detail on Br35H dataset.

Dataset	Rotation	Width shift range	Height shift range	Shear range	Zoom Range
Br35H	90 degree	0.2	0.2	0.2	0.2

Figure 3
Br35H-Augmented MRI images of Healthy and Tumor Class.

C) Proposed Model

With reference to the scope of this research work, a transfer learning-based EfficientNetB4 model with an adjustable custom learning rate is proposed for the classification of tumors in the brain with the addition of some hyperparameters, including a batch normalization layer that is responsible for providing stability and accelerating the training process by normalizing the input of each layer with a momentum of 0.99 and an epsilon of 0.001, which are responsible for accelerating convergence in the direction where gradients are consistent, preventing the denominator value from being zero, and providing numerical stability. Further, a dense layer with 256 neurons was added for more feature extraction with different regularizer, which are responsible for preventing overfitting by adding penalties to the loss function. For that, an l2 kernel regularizer with a value of 0.016 is applied, which applies penalties on layer parameters or weights, an l1 activity regularizer with a value of 0.006 is applied, which applies penalties on layer output; and an l1 bias regularizer with a value of 0.006 is applied, which applies penalties on bias terms, just to make better convergence of values and better output prediction with a relu activation function, which is responsible for learning complex functions from the data using a model and mitigating the vanishing gradient problem. Further, a dropout layer is applied with a ratio of 0.45, which is responsible for removing the complexity of the model to prevent overfitting by randomly dropping out a fraction of neurons during the training process, with an output layer as a fully connected layer. The model is trained on the Adamax optimizer with a learning rate of 0.001, and optimizers are especially useful for reducing the loss function by updating the model parameters by controlling the learning process based on the learning rate. The reason for choosing these hyperparameters is the best working of them and the nature of their producing better results than other hyperparameters.

The EfficientNet Model is introduced for the first time in 2019 by Tan and Le. These models are designed to achieve state-of-the-art performance in terms of accuracy by using fewer parameters and less computational power than other convolutional neural network models. The EfficientNet model has seven different variants, from B0 to B7. The main benefits of the EfficientNet model are that it can adjust the network dimension by using a compound scaling method that depends on depth, width, and resolution scaling. From all variants, we have chosen the EfficientNetB4 variant, which was based on a balance between model complexity and performance and offers a middle ground within the EfficientNet family, providing a good trade-off between computational efficiency and accuracy. On other convolutional models, only depth scaling is performed where the number of layers is increased for detail feature extraction, but in the EfficientNet Model, both width scaling and resolution scaling are also performed. In the case of width scaling, the number of channels should be adjustable when the dataset is too large or increased by the data augmentation approach, which means the number of features has increased and there will be a need for more channels to capture fine-grained features, while in resolution scaling, more and deeper pixels images with a bigger size are used to overcome accuracy and available resources in terms of memory and the number of possible operations.

The key depth scaling issue is how many layers to increase for custom data or for training a model on all datasets. Choosing how many channels to increase for different datasets is difficult, like width scaling. Thus, a compound scaling mechanism that uniformly scales depth, width, and resolution using a constant ratio by finding the best coefficients that maximize the accuracy of the proposed and variants of the EfficientNet Model under the available resources and a usable dataset could solve all these issues. Equations 1, 2, 3, and 4 indicate the highest accuracy possible with particular model depth, width, and resolution under different setups and ensure that it does not exceed preset memory or flops. Developing computationally efficient and fast-running models requires this limitation. The methods section discusses the compound coefficient factor, which drives the EfficientNet model. This compound coefficient automatically expands the model based on its width and height to meet use needs.

${\max }_{d,w,r}=Accuracy\left(\eta \left(d,w,r\right)\right)$ (1)

$Memory\left(\eta \right)\le T\arg et\_memory$ (2)

$Flops\left(\eta \right)\le T\arg et\_Flops$ (3)

Where: d: represent depth, w: represent width, r: represent resolution, and r: represent rate, on which decided to increase ratio.

In the proposed EfficientNetB4 model, the base structure is the same, which is EfficientNetB0, and it remains the same in all other variants, but the difference is in the depth, width, and resolution adjustable parameters that are based on the compound scaling method, where the coefficient (called the compound coefficient) value can be used to determine the variations in width, depth, and resolution.

$depth\left(d\right)={\alpha }^{\varnothing }$ (4)

$width\left(w\right)={\beta }^{\varnothing }$ (5)

$resolution\left(r\right)={\gamma }^{\varnothing }$ (6)

Where:$\varnothing$ Compound Coefficient, $\alpha$: Constant for depth, $\beta$: Constant for width, and $\gamma :$ Constant for resolution.

Here $\alpha ,\beta ,\gamma$ are the constant values, while $\varnothing$ is the compound coefficient defined by user that used to control the resources available for depth, width, and resolution scaling for a model.

In this research work, a pretrained EfficientNetB4 model is proposed; in pretrained models, the model is already trained on a larger dataset, such as ImageNet, and the target dataset can be fine-tuned on that model. This transfer learning approach could be very beneficial for the target dataset because of its pretrained nature on a larger dataset with learning of a wider range of features and patterns in images. It can then be fine-tuned for the target dataset, by which accuracy can be achieved with less computational resource usage. Below Figure 4, shows the base line Model EfficientNetB0 architecture, all other variants of the base model are the same except for the adjustment of the compound coefficient factor for depth, width, and resolution scaling.

Figure 4
EfficientNetB0 Base Model Architecture [³⁴].

EfficidentNetB4 is deeper than its based model and has three major components: the stem, which is a composite of convolutional layers for initial feature extraction from the images, and the base network, which is a series of repeated blocks, each of which has convolutional layers with batch normalization and activation functions. In Head, which a composite is of fully connected layers, those layers map the output to the target class. The actual size of the EfficientNetb4 model depends on the scaling coefficient used to customize the network for specific resources, but overall, 53 convolutional layers are included in the stem and base model parts. Below, Figure 5 shows proposed methodology workflow.

Figure 5
Proposed Methodology Workflow.

A custom callback is very useful when training the model and determining whether achievable matrices are satisfactory or not. If they are satisfactory, model training would be stopped, and a custom callback would return the weights of those epochs where monitored parameters such as accuracy or validation accuracy achieved the highest performance. Initially, callbacks monitor training accuracy and adjust the learning rate based on a threshold value that is adjusted to 0.9; once training accuracy reaches the desired threshold, callbacks start to monitor validation loss and adjust the learning rate to decrease validation loss. In our model, patience is set to 1, and after 1 epoch if there is no improvement, the learning rate will be reduced. Training will be stopped. After three consecutive adjustments to the learning rate with no improvements, we set a 0.9 value for the threshold, and once the training accuracy reaches the threshold, callback adjust the learning rate for the validation loss. When the learning rate is adjusted, the new learning rate is equal to 0.5x then adjusted learning rate, and by default, a 0.001 learning rate is set. Training starts from 0 epoch and after every 5-epoch system halt and ask to further run more epochs, and in the end return weight of high achieving performance epoch return and system classify tumor on that achieving metrics, and table 5 shows the hyperparameters detail used to train the proposed model.

D) Pseudo Algorithm

Step1: Acquired the Br35H dataset from Kaggle.

Step2: Generate the Augmented dataset for increasing number of images.

Step3: Split the dataset into training, validation and testing

Step3: Dataset preprocessed using below steps

3.1 Dataset Resizing

3.2 Labels converted to NumPy arrays

3.3 One hot encoding applied

3.4 Normalize the input images

Step4: Import the EfficientNetB4 model and applied modification

Step5: Proposed model trained on Br35H and Augmented-Br35H dataset

Step6: performance of the model evaluated regarding different performance evaluation parameters.

Table 5
Hyperparameters used to train proposed EfficientNetB4 model.

Dataset	Image size	Batch size	Epoch	Patience	Threshold	Optimizer	Activation function	Loss function
Augmented-Br35H data	300 × 300	20	30	1	0.9	Adamax	Sigmoid	Binary-cross entropy

RESULTS AND DISSCUSSIONS

This section demonstrates the performance of an adjustable learning rate-based, pre-trained EfficienetNetb4 model on the Br35h dataset along with the Br35h-Augumented dataset. The performance of the model can be evaluated in terms of accuracy [³⁵], sensitivity [³⁶], specificity [³⁷], precision, F1-score [³⁸], false-omission rate [³⁹], negative predictive rate [⁴⁰], miss classification rate [⁴¹], false negative rate, and false positive rate.

$Accuracy=\frac{\left(TN+TP\right)}{\left(TN+TP+FN+FP\right)}\times 100$ (7)

$MissClassificationRate\left(MCR\right)=\frac{\left(FP+FN\right)}{\left(TN+FN+FP+TP\right)}\times 100$ (8)

$Sensitivity=\frac{TP}{\left(TP+FN\right)}\times 100$ (9)

$Specificity=\frac{TN}{\left(TN+FP\right)}\times 100$ (10)

$F1-Score=\frac{2TP}{\left(2TP+FP+FN\right)}\times 100$ (11)

$FalseOmissionRate\left(FOR\right)=\frac{FN}{\left(FN+TN\right)}\times 100$ (12)

$\mathrm{Pr}ecison\left(PPV\right)=\frac{TP}{\left(FP+TP\right)}\times 100$ (13)

$Negative\mathrm{Pr}edictiveValue\left(NPV\right)=\frac{TN}{\left(TN+FN\right)}\times 100$ (14)

$FalseNegativeRate\left(FNR\right)=\frac{FN}{\left(FN+TP\right)}\times 100$ (15)

$FalsePositiveRate\left(FPR\right)=\frac{FP}{\left(FP+TN\right)}\times 100$ (16)

True Negative (TN) is the value or outcome where the model correctly predicts the negative class; True Positive (TP) is the value of the outcome where the model correctly predicts the positive class; just like it, False Negative (FN) and False Positive (FP) are the outcomes where the model wrongly predicts the negative and positive classes, respectively. Table 6 shows the confusion matrix values for Br35h, and Br35h-augmented validating data, while Table 7 and 8 show the statistical evaluation of the proposed model on the Br35h and Br35h augmented datasets regarding different statistical performance evaluation parameters.

Table 6
Br35H, and Augmented Br35H dataset confusion Matrix (TP, FP, FN, and TP).

CNN Model	Dataset	True Negative (TN)	False Positive (FP)	False Negative (FN)	True Positive (TP)
EfficientNetB4	Br35h	149	0	1	150
EfficientNetB4	Br35h-Augumented	391	1	0	390

Table 7
Performance evaluation regarding Accuracy MCR, Sensitivity, Specificity, and Precision.

CNN Architecture	Dataset	Accuracy	MCR	Sensitivity	Specificity	Precision
EfficientNetB4	Br35h	99.67	0.33	99.34	100.00	100.00
EfficientNetB4	Br35h-Augumented	99.87	0.13	100.00	99.74	99.74

Table 8
Performance evaluation regarding NPV, FOR, F1-Score, FNR, and FPR.

CNN Architecture	Dataset	NPV	FOR	F1-Score	FNR	FPR
EfficientNetB4	Br35h	99.33	0.67	99.67	0.00	0.26
EfficientNetB4	Br35h-Augumented	100.00	0.67	99.67	0.66	0.00

Table 7, and 8 shows the evaluation of the performance of the proposed pretrained EfficientNetb4 model on two different dataset samples with different evaluation parameters. With the Br35h dataset, the proposed model achieved accuracy of 99.66% with a miss classification rate (MCR) of 0.33%, further sensitivity, specificity, and precision values are 99.33%, 99.97%, and 99.93%, respectively. The negative predictive value (NPV), false omission rate (FOR), and F1-score obtained results are 99.33%, 0.66%, and 99.66%. The FNR and FPR values achieved with values of 0 and 0.26%, respectively. After applying the augmentation technique to the same dataset with different parameters, the performance of the proposed architecture increased, achieving an accuracy of 99.87% with a miss classification rate of 0.12%. The proposed model also achieved a sensitivity of 99.97%, a specificity of 99.74%, and a precision of 99.74%. On augmented data, the negative predictive value (NPV), false omission rate (FOR), and F1-score values are 100%, 0.67%, and 99.67%, respectively. The FNR and FPR values achieved with values of 0.66% and 0%, respectively. In the above table, it is clearly shown that the proposed model performed well on an augmented dataset with improvements in accuracy, MCR, sensitivity, and NPV. The Figure 6 shows the graphical representation of proposed model with different performance evaluation parameters.

Figure 6
Performance evaluation of EfficientNetB4 Model.

When we talked about how well the proposed model classifies different classes in terms of brain tumor detection, a confusion matrix can be beneficial because it can quantify how well each class can be classified. The proposed pretrained EfficientNetB4 model showed excellent results for the binary classification of brain tumors in both the augmented-Br35h and Br35h datasets. The Figure 7 and Figure 8 shows the confusion matrix for the proposed model.

Figure 7
Confusion Matrix of Br35H

Figure 8
Confusion Matrix of Br35H-Augmented

In first confusion matrix, only 1 value could be false predictive that belong to tumor class, while in second confusion matric 1 value predicted wrong that belongs to healthy class. Below Figure 9 and Figure 10 showed graph of how training and validation loss and accuracy curves maintained by proposed model in both used datasets, and showed on which epoch values system achieved best performance.

Figure 9
Training and Validation Loss, and Accuracy for Br35H.

Figure 10
Training and Validation Loss, and Accuracy for Br35H-Augmented.

In order to measure the performance of the proposed model, we also measured the performance of the proposed model on KFold cross validation [⁴²-⁴³], which is providing averaging results from multiple training and validation splits and providing a more comprehensive view of how the model works. For this, we applied 5Fold cross validation to the Br35H-augumented dataset because the model produced better results on augmented data, and Table 9 shows the results of 5Fold cross validation.

Table 9
Cross Fold validation of proposed model

Fold	Train Loss	Train Accuracy	Val Loss	Val Accuracy
Fold-1	0.0053	98.96	0.37	97.39
Fold-2	0.0059	98.95	0.29	98.31
Fold-3	0.0036	99.19	0.19	98.37
Fold-4	0.0032	99.22	0.16	99.09
Fold-5	0.0027	99.64	0.11	99.87
Final Performance	0.00414	99.192	0.224	98.606

The 5-fold cross-validation model achieved overall results of 98.6% validation accuracy with a loss of 0.224%. Table 10 shows the comparison of the proposed pre-trained EfficientNetB4 model with existing state-of-the art methods. The results in terms of accuracy and miss classification rate showed that the proposed pre-trained EfficientNetB4 model is more accurate and efficient as compared to previous models. The below figure 11 show the testing results of the dataset taken from Kaggle.

Figure 11
Testing different images taken from Kaggle

Table 10
Comparison of Proposed model with recent state of the art model.

References	Methods	Dataset	Accuracy	MCR
[20]	MobileNetV2	Kaggle (2513 cancerous, 2087 healthy)	97%	3%
[25]	EADL-BTMIC	N. A	99.06%	0.94%
[26]	Modified VGG-16	10153MRI images	99.5%	0.50%
[27]	MBINet Model using (Self-ONN)	1320 RMB images	98%	2%
[29]	Hybrid googleNet, alexNet	Online publicly available dataset	99.51%	0.49%
[3]	Hybrid transfer learning	Br35h dataset	99.51%	0.49%
[30]	Generic CNN	Msoud (combinantion of Sartaj, Figshare, and Br35h dataset) 7023 images	97.12%	2.88%
[31]	ResNet 18	3064 MRI images from publicly available	99.56%	0.44%
[32]	ViT	NA	>90%	<10%
[33]	CNN Log-polar	Formed by itself	96%	4%
Khushi et al.	Proposed Model	Br35H	99.67%	0.33%
Khushi et al.	Proposed Model	Br35H-Augmented	99.87%	0.13%

In above comparison table 10, we have compared the results of or proposed model regarding Br35H and Augmented Br35H dataset which shows that our model produced better results regarding other state of the art approaches regarding accuracy, and loss which means it’s also produced better results regarding other performance evaluation parameters and due to unavailability of other matrices in different state of the art approaches, we just compares results regarding accuracy and loss.

CONCLUSION

This research work was conducted with the intention of detecting tumors in the brain, which was done by using a pre-trained EfficientNetB4 model and applying custom callbacks for the adjustment of learning rate according to validation accuracy loss. The main aim was to make a system that can detect brain tumors very early and very efficiently and accurately, compared to other state-of-the-art available models. Model implemented on a Br35h publicly available dataset, with enhancement of data by applying some extra steps during the data augmentation approach. The proposed model achieved 99.67% accuracy with a miss classification rate of 0.33% for br35h, while this model achieved 99.87% accuracy with a 0.13% miss classification rate on the br35h-augmented dataset. The proposed pretrained EfficientNetB4 model achieved higher accuracy with a lower miss classification rate as compared to other state-of-the-art architectures.

LIMITATIONS AND FUTURE WORK

This research is proposed for the binary classification of tumors in the brain, which classify either tumors exist or do not on the br35h dataset. For further use, this proposed model will be applied for the multiclass classification of tumors in the brain with different available datasets. Further, this proposed methodology will be tested in other medical domains also with more variants of EfficientNetB0 Models.

REFERENCES

• ¹ Van Meir EG, Hadjipanayis CG, Norden AD, Shu HK, Wen PY, Olson JJ. Exciting New Advances in Neuro-Oncology: The Avenue to a Cure for Malignant Glioma. CA Cancer J Clin. 2010 May 1;60(3):166-93.
• ² Bush NAO, Chang SM, Berger MS. Current and future strategies for treatment of glioma. Neurosurg Rev. 2016 Apr 16;40(1):1-14.
• ³ Amran GA, Alsharam MS, Blajam AOA, Hasan AA, Alfaifi MY, Amran MH, et al. Brain Tumor Classification and Detection Using Hybrid Deep Tumor Network. Electronics. 2022 Oct 25;11(21):3457.
• ⁴ Jagannathan J, Kanter AS, Sheehan JP, Jane JA, Laws ER. Benign Brain Tumors: Sellar/Parasellar Tumors. Neurol. Clin. 2007 Nov;25(4):1231-49.
• ⁵ Dutoit V, Migliorini D, Dietrich PY, Paul Richard Walker. Immunotherapy of Malignant Tumors in the Brain: How Different from Other Sites? Front Oncol. 2016 Dec 7;6.
• ⁶ Villarraga-Gómez H, Herazo EL, Smith ST. X-ray computed tomography: from medical imaging to dimensional metrology. Precision Engineering [Internet]. 2019 Nov;60:544-69. Available from: https://www.sciencedirect.com/science/article/pii/S0141635919300376
  » https://www.sciencedirect.com/science/article/pii/S0141635919300376
• ⁷ Tandel GS, Biswas M, Kakde OG, Tiwari A, Suri HS, Turk M, et al. A Review on a Deep Learning Perspective in Brain Cancer Classification. Cancers. 2019 Jan 18;11(1):111.
• ⁸ Gorla P, Agrawal A. Hybrid approach for brain tumor detection and classification in magnetic resonance images. 2015 Nov 1.
• ⁹ Ouyang W, Zeng X, Wang X, Qiu S, Luo P, Tian Y, et al. DeepID-Net: Object Detection with Deformable Part Based Convolutional Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence [Internet]. 2017 Jul 1 [cited 2020 Nov 8];39(7):1320-34. Available from: https://wang-zhe.me/welcome_files/papers/ouyangZWpami16.pdf
  » https://wang-zhe.me/welcome_files/papers/ouyangZWpami16.pdf
• ¹⁰ Darko Zikic, Glocker B, Ender Konukoglu, Criminisi A, Çağatay Demiralp, Shotton J, et al. Decision Forests for Tissue-Specific Segmentation of High-Grade Gliomas in Multi-channel MR. 2012 Oct 1;369-76.
• ¹¹ Elisabetta Binaghi, Massimo Omodei, Pedoia V, Balbi S, Lattanzi D, Monti E. Automatic Segmentation of MR Brain Tumor Images using Support Vector Machine in Combination with Graph Cut. 2014 Jan.
• ¹² Bahadure NB, Ray AK, Thethi HP. Comparative Approach of MRI-Based Brain Tumor Segmentation and Classification Using Genetic Algorithm. J. Digit. Imaging. 2018 Jan 17;31(4):477-89.
• ¹³ Revannath Dnyandeo Nikam, Lee J, Choi W, Banerjee W, Kwak M, Yadav M, et al. Ionic Sieving Through One‐Atom‐Thick 2D Material Enables Analog Nonvolatile Memory for Neuromorphic Computing. Small. 2021 Oct 1;17(44).
• ¹⁴ Khalid S, Khalil T, Nasreen S. A survey of feature selection and feature extraction techniques in machine learning. 2014 Science and Information Conference. 2014 Aug.
• ¹⁵ Erickson BJ, Korfiatis P, Akkus Z, Kline TL. Machine Learning for Medical Imaging. RadioGraphics. 2017 Mar;37(2):505-15.
• ¹⁶ Harish P, Baskar S. MRI based detection and classification of brain tumor using enhanced faster R-CNN and Alex Net model. Materials Today: Proceedings. 2020 Dec.
• ¹⁷ He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2016 Jun;770-8. Available from: https://ieeexplore.ieee.org/document/7780459
  » https://ieeexplore.ieee.org/document/7780459
• ¹⁸ Zhang X, Zhou X, Lin M, Sun J. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. 2018 IEEE/CVF Conf Comput Vis Pattern Recognit. 2018 Jun 1;6848-56.
• ¹⁹ Saleem S, Amin J, Sharif M, Anjum MA, Iqbal M, Wang SH. A deep network designed for segmentation and classification of leukemia using fusion of the transfer learning models. Complex & Intelligent Systems. 2021 Jul 28.
• ²⁰ Mohammad Monirujjaman Khan, Atiyea Sharmeen Omee, Tahia Tazin, Almalki FA, Maha Aljohani, Haneen Algethami. A Novel Approach to Predict Brain Cancerous Tumor Using Transfer Learning. Computational and Mathematical Methods in Medicine. 2022 Jun 20;2022:1-9.
• ²¹ Tahir A, Asif M, Ahmad MB, Mahmood T, Khan MA, Ali M. Brain Tumor Detection using Decision-Based Fusion Empowered with Fuzzy Logic. Sajid M, editor. Mathematical Problems in Engineering. 2022 Aug 21;2022:1-13.
• ²² Saravanan S, Kumar VV, Sarveshwaran V, Indirajithu A, Elangovan D, Allayear SM. Computational and Mathematical Methods in Medicine Glioma Brain Tumor Detection and Classification Using Convolutional Neural Network. Ravi V, editor. Computational and Mathematical Methods in Medicine. 2022 Oct 14;2022:1-12.
• ²³ Siddiqi MH, Alsayat A, Alhwaiti Y, Azad M, Alruwaili M, Alanazi S, et al. A Precise Medical Imaging Approach for Brain MRI Image Classification. Ahmad S, editor. Comput. Intell Neurosci. 2022 May 2;2022:1-15.
• ²⁴ Sangeetha SKB, Muthukumaran V, Deeba K, Rajadurai H, Maheshwari V, Dalu GT. Multiconvolutional Transfer Learning for 3D Brain Tumor Magnetic Resonance Images. Comput. Intell. Neurosci. [Internet]. 2022 Aug 23 [cited 2024 Jan 19];2022:e8722476. Available from: https://www.hindawi.com/journals/cin/2022/8722476/.
  » https://www.hindawi.com/journals/cin/2022/8722476
• ²⁵ Manar Ahmed Hamza, Hanan Abdullah Mengash, Alotaibi SS, Haj B, Ayman Yafoz, Fahd Althukair, et al. Optimal and Efficient Deep Learning Model for Brain Tumor Magnetic Resonance Imaging Classification and Analysis. 2022 Aug 8;12(15):7953-3.
• ²⁶ Ahmed Wasif Reza, Muhammad Sazzad Hossain, Moonwar Al Wardiful, Farzana M, Ahmad S, Alam F, et al. A CNN-Based Strategy to Classify MRI-Based Brain Tumors Using Deep Convolutional Network. Applied Sciences. 2022 Dec 27;13(1):312-2.
• ²⁷ Hossain A, Islam MT, Abdul Rahim SK, Rahman MA, Rahman T, Arshad H, et al. A Lightweight Deep Learning Based Microwave Brain Image Network Model for Brain Tumor Classification Using Reconstructed Microwave Brain (RMB) Images. Biosensors. 2023 Feb 7;13(2):238.
• ²⁸ Tummala S, Kadry S, Bukhari SAC, Rauf HT. Classification of Brain Tumor from Magnetic Resonance Imaging Using Vision Transformers Ensembling. Current Oncology. 2022 Oct 7;29(10):7498-511.
• ²⁹ Samee NA, Mahmoud NF, Atteia G, Abdallah HA, Alabdulhafith M, Al-Gaashani MSAM, et al. Classification Framework for Medical Diagnosis of Brain Tumor with an Effective Hybrid Transfer Learning Model. Diagnostics. 2022 Oct 20;12(10):2541.
• ³⁰ Gómez-Guzmán MA, Jiménez-Beristaín L, García-Guerrero EE, López-Bonilla OR, Tamayo-Perez UJ, Esqueda-Elizondo JJ, et al. Classifying Brain Tumors on Magnetic Resonance Imaging by Using Convolutional Neural Networks. Electronics [Internet]. 2023 Jan 1;12(4):955. Available from: https://www.mdpi.com/2079-9292/12/4/955
  » https://www.mdpi.com/2079-9292/12/4/955
• ³¹ Usmani IA, Qadri MT, Zia R, Alrayes FS, Saidani O, Dashtipour K. Interactive Effect of Learning Rate and Batch Size to Implement Transfer Learning for Brain Tumor Classification. Electronics [Internet]. 2023 Jan 1;12(4):964. Available from: https://www.mdpi.com/2079-9292/12/4/96
  » https://www.mdpi.com/2079-9292/12/4/96
• ³² Uysal F, Erkan M. Multiclass Classification of Brain Tumors with Various Deep Learning Models. Eng Proc 2022, Vol 27, Page 30 [Internet]. 2022 Nov 1 [cited 2024 Aug 10];27(1):30. Available from: https://www.mdpi.com/2673-4591/27/1/30/htm
  » https://www.mdpi.com/2673-4591/27/1/30/htm
• ³³ Jibon FA, Khandaker MU, Miraz MH, Thakur H, Rabby F, Tamam N, et al. Cancerous and Non-Cancerous Brain MRI Classification Method Based on Convolutional Neural Network and Log-Polar Transformation. Healthcare. 2022 Sep 19;10(9):1801.
• ³⁴ Khan T, Gyu Ho Choi, Lee S. EFFNet-CA: An Efficient Driver Distraction Detection Based on Multiscale Features Extractions and Channel Attention Mechanism. 2023 Apr 8;23(8):3835-5.
• ³⁵ Mao Q, Zhao S, Ren L, Li Z, Tong D, Yuan X, et al. Intelligent immune clonal optimization algorithm for pulmonary nodule classification. Math Biosci Eng. 2021;18(4):4146-61.
• ³⁶ Gao Y, Song F, Zhang P, Liu J, Cui JJ, Ma Y, et al. Improving the Subtype Classification of Non-small Cell Lung Cancer by Elastic Deformation Based Machine Learning. 2021 May 7.
• ³⁷ Lai KD, Nguyen TT, Le TH. Detection of Lung Nodules on CT Images based on the Convolutional Neural Network with Attention Mechanism. Annals of Emerging Technologies in Computing. 2021 Apr 1;5(2):78-89.
• ³⁸ Zhang D, Wang J, Zhao X. Estimating the Uncertainty of Average F1 Scores. Proceedings of the 2015 International Conference on The Theory of Information Retrieval. 2015 Sep 27.
• ³⁹ Kost GJ. The Impact of Increasing Prevalence, False Omissions, and Diagnostic Uncertainty on Coronavirus Disease 2019 (COVID-19) Test Performance. Archives of Pathology & Laboratory Medicine. 2021 Mar 8.
• ⁴⁰ Salvo G, Ramírez PT, Levenback C, Munsell MF, Euscher ED, Soliman PT, et al. Sensitivity and negative predictive value for sentinel lymph node biopsy in women with early-stage cervical cancer. Gynecologic Oncology. 2017 Apr 1;145(1):96-101.
• ⁴¹ Wei Zhong Liu, White AP, Thompson SG, Bramer MA. Techniques for dealing with missing values in classification. Lecture notes in computer science. 1997 Jan 1;527-36.
• ⁴² Ozdemir C, Dogan Y. Advancing brain tumor classification through MTAP model: an innovative approach in medical diagnostics. Med Biol Eng Comput. 2024 Mar 14.
• ⁴³ Ozdemir C. Classification of Brain Tumors from MR Images Using a New CNN Architecture. Traitement du signal/TS Traitement du signal. 2023 Apr 30;40(2):611-8.
• ⁴⁴ Atasever S, Azgınoglu N, Terzı DS, Terzı R. A comprehensive survey of deep learning research on medical image analysis with focus on transfer learning. Clin. Imaging. 2022 Nov.
• ⁴⁵ Khaliki MZ, Başarslan MS. Brain tumor detection from images and comparison with transfer learning methods and 3-layer CNN. Scientific Reports [Internet]. 2024 Feb 1;14(1):2664. Available from: https://www.nature.com/articles/s41598-024-52823-9
  » https://www.nature.com/articles/s41598-024-52823-9
• ⁴⁶ Br35H: Brain Tumor Detection 2020 [Internet]. www.kaggle.com. Available from: https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection
  » https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection

• Funding:
  This research received no external funding.

Edited by

• Editor-in-Chief:
  Alexandre Rasi Aoki

• Associate Editor:
  Raja Soosaimarian Peter Raj

Publication Dates

• Publication in this collection
  08 Nov 2024
• Date of issue
  2024

History

• Received
  19 Nov 2023
• Accepted
  29 July 2024