**3. Brain MRI Dataset and Results**

This study validates the suggested paradigm using an online collection of brain MRI images [39]. The dataset for this study was obtained from the Kaggle website [39]. It contains three tumor classes (glioma, pituitary, and meningioma) and one class of no tumor. It has 2870 brain MRI images in total. Additionally, 80% of the data of each class were utilized for the training of the models. The remaining 20% of the data were used to test the trained models. The brain MRI images and percentage distribution of images per class are shown in Figure 2.

**Figure 2.** (**a**) The brain MRI images of each class; (**b**) the percentage distribution of MRI images per class.

In this work, MATLAB 2021 was utilized for training the models in the 64-bit Windows 11 operating system (core i7, 11th generation, 32 GB RAM, NVIDIA GeForce GTX 1060, and 1 TB SSD). In addition, the classification accuracy was used as a comparison metric for the various trained models (SVM, tree, Naïve Bayes, K-NN, ensemble, and NN). The results of the KAZE- and SURF-trained models are presented in Figure 3.

**Figure 3.** The comparison of various machine learning models for SURF and KAZE features.

It is evident from Figure 3 that the SVM model trained with SURF and KAZE features shows accuracies of 93.4% and 93.7%, respectively, which are the highest among all methods. Therefore, it may be fruitful to concatenate the features of SURF and KAZE to determine the model's performance in classifying brain MRI images. Furthermore, the confusion matrixes of the SURF-, KAZE-, and SURF + KAZE- (hybrid) trained SVM models are shown in Figure 4.

*Life* **2022**, *12*, 1084


(**a**) (**b**)

243

**Figure 4.** Confusion matrixes of various models: (**a**) SURF-trained SVM; (**b**) KAZE-trained SVM; (**c**) SURF + KAZE (hybrid)-trained SVM (proposed model).

(**c**)

The SVM model trained with concatenation features shows the highest accuracy of 95.33%, almost 2% higher than the SVM model trained with SURF features. Therefore, the proposed SURF + KAZE-trained SVM model has true positive rates (TPRs) of 97.75% and 98.42% for the glioma and pituitary tumor classes. Furthermore, the proposed model correctly classifies 19 more MRI brain images for the no tumor class than the KAZE-trained model. Similarly, 36 more brain MRI images are correctly classified for the meningioma tumor class compared to the SURF-trained model. Finally, the proposed model is compared with the pre-trained deep-feature-trained SVM model presented by Kang et al. [25]. The comparison results of various SVM models are presented in Figure 5.

(**b**)

**Figure 5.** Comparison of SVM model trained with deep features with the proposed model: (**a**) accuracy comparison; (**b**) accuracy and computational complexity.

For further validation of the proposed approach, another public dataset is utilized [40]. The dataset contains a total of 3064 brain MRI scans. Further details about the dataset are shown in Figure 6. The classification result of the new dataset is presented in Figure 7.

**Figure 6.** The percentage distribution per class of brain MRI dataset [40].

**Figure 7.** Confusion matrix of the proposed model for new dataset [40].
