The general health of the people and livestock and nature, in general, is considered the foremost wealth component of any nation. Thus, improving health and controlling diseases are crucial factors in the sustenance and progress of the world. Early identification of diseases is crucial in disease control. Therefore, the rapid and accurate diagnosis is of foremost significance. There has been a steady growth in the medical instrumentation field in the past and present centuries. With the advent of computers, accurate interpretation of data analysis and measurements has led to vast improvement. Computer-aided analytical tools have become a great help to medical experts in decision-making. Computer-aided diagnosis is a fast-growing research area. Medical image processing (MIP) is one of the important techniques in diagnosis, where classification is a very important process to classify the disease, whether benign or malignant.

Successful treatment of diseases depends largely on early detection and accurate assessment of the state of the disease. Disease diagnosis is an important process in the maintenance of global well-being. Modern diagnosis involves measurement, analysis, and decision-making. Computers play a critical role in medical decision support. Among the fatal diseases, brain cancer is particularly difficult because it is not usually detected until it is too late for prognosis. MRI is focused very specifically because of its effectiveness and harmlessness. The brain tumor classification is essential for the treatment plan and further the assessment process of the tumor.

In this study, we have proposed and investigated seven different classifiers to efficiently classify the MRI brain tumor images into benign or malignant types. The deep learning (DL)-based neural network architecture is designed to group each pixel into a voxel of interest, capturing the required information by choosing the input to be applied in the deep network. The choice of input contributes to avoiding overfitting and reducing computational complexity significantly. Selecting the local precision and global spatial intensity features are considered as most important in this study. The experimental results are validated using the evaluation metrics such as sensitivity, specificity, and accuracy. Among the seven classifiers, the combined feature and image-based classifier (CFIC) performs significantly better than the existing classification methods.

The remaining sections are organized as follows: related works using various methods are discussed in image classification, proposed work of the various classifiers are presented, results and discussions of the proposed methods are compared, and finally, the conclusion and future works are discussed.

Various methods are available for brain tumor classification. Among various methods, neural network-based methods, convolutional neural network (CNN) methods, and DL methods are widely used.

The study has investigated the use of the deep features extracted from the CNNs that are pretrained in the prediction of survival time. It provides further evidence for domain-specific fine-tuning to improve its performance. Standard dataset is available in the internet. It has an accuracy of approximately 81% in the case of leave-one-out-based cross-validation (Ahmed et al., 2017).

The study has proposed a hybrid method for classifying the tissues of the brain tumor image. In this technique, the system employs a genetic algorithm (GA) for feature extraction and an support vector machine (SVM) for classification purposes. The features are further compared with the stored features, and the method is used to capture the images and their visual contents. This method signifies the raw image to simplify decision-making in a very focused form. The choice of the features composed is very difficult in the classification methods that are duly solved using the GA. The features with the SVM have been used to classify the tumor as normal or abnormal. If the tumor is detected based on mean, mode, and median, then the tumor is classified as either meningioma or pituitary tumor. The performance of the algorithm is assessed based on the images containing the brain tumor (Bangare et al., 2017).

The study has found the solutions for classifying MRI brain tumors using the hybrid technique. This method gives a second opinion for the physicians to succeed in the treatment process. It works only on the specific type of tumor with limited image datasets (Mohan and Subashini, 2018).

The study has proposed a simple three-step algorithm for identifying brain tumors in MRI images. The process consists of identifying patients with brain tumor, automatically selecting all the abnormal slices of such patients, thereby segmenting, and recognizing the tumor. The features are extracted using the discrete wavelet transform (DWT) on several normalized images, which are further classified using the SVM algorithm for the first process. The same procedure is followed for the second process. The random forest (RF) classification algorithm is used instead of SVM. Another 400 subjects are divided into a ratio of 3:1 without any overlap. All the T2-weighted slices were configured as one single image-based classification. It was equipped with a new and unique contralateral approach, with patch thresholds for the segmentation process. This patch threshold does not require training sets or templates used for the investigations of segmentation. The tumors are segmented with high accuracy compared with the other classification methods (Gupta et al., 2017).

The study has found that deep learning is a new approach in the machine learning field, which is applied in many complex application areas. This DL approach, combined with DWT and principal component analysis (PCA), leads to better brain tumor classification performance. The lack of a real-time dataset for getting better results is less and less convenient in time (Mohsen et al., 2018).

The study has presented the deep feature and machine learning-based MRI brain tumor image classification method to classify the MRI brain tumor images into benign and malignant. The different CNN classifier models are compared with the deep feature and machine learning classifier models, and better result is achieved (Kang et al., 2021). The local constraint-based convolutional dictionary learning is the new method to classify brain tumors into normal or abnormal types. This supervised classification method uses the REMBRANDT datasets for classification purposes. This method helps to classify the tumor significantly better than the various existing methods (Gu et al., 2021).

The study has presented the modified deep convolutional neural network (DCNN) to efficiently classify brain tumor images into the benign and malignant. The MRI brain images of T1, T2, and T2 fLuid-attenuated inversion recovery (FLAIR) of 220 cases are used to classify using this method with a better accuracy rate (Hemanth et al., 2019). The study has discussed the DL-based opposition crow search (DL-OCS) method to efficiently classify MRI brain tumor images. In this technique, the optimal feature selection is considered as most important for the classification process. The DL-OCS method performs better in terms of sensitivity (86.45%), specificity (100%), and accuracy (95.22%) (Raj et al., 2020).

The study has discovered various key research fields in medical image analysis using machine learning algorithms. The machine learning algorithms are helped to maintain the medical records in digital formats, increasing over time. The advantages of digitizing the medical images are it is accurate, up-to-date, quickly accessible, more reliable, easy to share, enhances privacy and security, reduces costs, etc. The important issue is the lack of datasets for training in medical image analysis. Furthermore, rare diseases are difficult to be spotted with reliable accuracy due to data imbalance or lack of training samples (Ker et al., 2018).

The study has presented the new progressive growing generative adversarial network (PGGAN) method based on CNN, which is an augmentation method that allows the maximization of the learning optimization of the network. The performance of the PGGAN method in terms of accuracy has been increased considerably (Khan et al., 2020). The study has presented a CNN with image processing and data augmentation methods to classify the MRI brain tumor images into cancerous or non-cancerous. The experiments have been carried out with very small datasets, compared with various existing transfer learning methods, which show better performance (Ahmed et al., 2021). The study has presented the new CNN approach for T1-weighted MRI brain tumor image classification. Then network was tested using the 10-fold validation techniques. It gives better accuracy than the traditional classification methods (Badza and Barjaktarovic, 2020).

The study has discussed two different classifiers based on the neural networks methods for MRI brain tumor images. This method consists of three stages: feature extraction, selection, and classification. The features are extracted using DWT, and the most relevant features are selected using PCA methods. The classification process is the final stage based on supervised learning techniques. They are feed-forward artificial neural networks and backpropagation neural network-based methods, and all these methods help classify the tumor into normal or abnormal MRI based on the brain tumor images (Kharat et al., 2012).

The study has proposed a method for segmenting and classifying the MRI brain tumor. The input MRI image is preprocessed by using the Ostu method for choosing the threshold. Furthermore, the tumor is detected using the K-means clustering. The image features are extracted using DWT and the Gabor wavelet. The PCA was used to reduce the feature set. Finally, the SVM classifier is used to classify the tumor image as benign or malignant (Mathew et al., 2017).

The study has found that many neurological diseases are identified using MRI images rather than other images. Early detection of brain tumors saves human life. Deep neural network (DNN)-based algorithms are widely used for various applications, especially for brain tumor segmentation and classification tasks. In this study, the deep wavelet autoencoder combined with DNN helps to reduce the feature dataset and gives better performance (Mallick et al., 2019).

Classification is a crucial task in data mining and has many applications. There are several applications in computer vision for this challenging task. Data belonging to different types can be classified using this approach to categorize every item within a dataset into a group or class that is predetermined. One set of the training records is used for the problem definition so that every record is named with a class value extracted from a set of k varied discrete values indexed by {1,., k}. These training data are utilized to formulate the classification models relating to features in an underlying record to one class of labels. For example, this training model can be utilized for forecasting the record to a class label when the class of an instance is unknown (Sravani et al., 2014).

The MRI brain tumor image classification is a process that plays a vital role in identifying and classifying dangerous diseases, either benign or malignant. Research has been conducted to diagnose brain tumours based on medical images. Two complementary qualities are required for strong classification performance, which are the descriptiveness and discriminativeness of the extracted features. In classification, machine learning is critical because of its wide range of approaches and suitability to a given problem.

The Kaggle Brain Tumor Detection 2020 dataset has been used to train and test the proposed classifiers. The proposed seven architectures for the classification of MRI brain tumor images are described in the “Proposed Architectures for Brain Tumor Classification” section. This section tabulates and discusses comparison of the performance of the proposed architectures.

The performance evaluation metrics are as follows:

where TP_p is the malignant image correctly predicted as malignant, TN_p is the benign image correctly predicted as benign, FP_p is the benign image incorrectly predicted as malignant, and FN_p is the malignant image incorrectly predicted as benign.

For the classification problem, seven different combinational architectures are investigated, each showing improvements in terms of one or more performance metrics. The AIFC is performing better than the SIFC. The ASIFC is performing better than both AIFC and SIFC. The implication is that the border information is last in a segmented image. The AIC is performing better than the SIC. The ASIC is better than both AIC and SIC. However, the CFIC is performing better when combined with ASIFC and ASIC in the accuracy. Finally, the CFIC method is the best among the proposed classification methods. The proposed DNN-based CFIC method gives significantly better results in terms of sensitivity, specificity, and accuracy with 98.86, 97.14, and 98.97%, respectively, compared with the next best existing techniques, namely CNN-ML and DTL with 97.72, 96.99, and 98.83%, respectively. The following are the directions for future work: (a) The algorithms developed should be incorporated into the software used by physicians and (b) the methods and techniques propounded in this study can only be applied to gray images. Further work could employ color images for the same problems.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.

AV, SM, and GM contributed in problem formulation, implementation and results. KK, JS, VV, and VS contributed in methodology identification and techniques to project results. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.