Feature Fusion Based Effective Brain Tumor Detection Approach Using MRI

Early identification of brain tumors greatly influences the clinical diagnosis process of a brain tumor patient. Therefore, this study suggests a brain tumor detection approach that merges deep and shallow features extracted from the brain MRI images in order to distinguish between non-tumor and tumor classes. We combine some pre-trained deep CNN architectures and the concept of transfer learning in our proposed framework to obtain high-level features from magnetic resonance images. Following the extraction, a Support Vector Machine classifier with radial basis function was used to evaluate the deep features. A deep feature vector is then created by combining the best three deep features that perform well on the SVM classifier. Even though deep features are crucial for classification, as the network becomes deeper, some low-level features might be lost. Therefore, a shallow network was intended to learn low-level information from the brain MRI. Deep and shallow features are then merged to compensate for the information loss. The fused feature vector is then employed, in order to train a support vector machine classifier. The experimental results were obtained on a publicly available dataset. Our proposed framework has achieved a high accuracy of 92.48% (with a precision of 93.64%, recall of 94.55%, and f1-score of 93.97%). The results also showed that utilizing this feature fusion enhances the performance of the classification framework and these results ensure the hypothesis that features fusion enables the compensation of low-level information lost. Moreover, our classification approach outperformed others when compared to state-of-the-art studies.