Fusion of Handcrafted and Deep-Learning Features for Brain Tumor Detection and Classification Using T1-Weighted Magnetic Resonance Images

The proliferation of irregular brain cells results in a condition known as a brain tumor (BT). Assessing a patient’s likelihood of survival can be difficult given the variety of tumor shapes and low incidence rate. While detecting cancer manually can be challenging and time-consuming, there is a potential for false-positive results. MRI is a crucial method for locating cancer. Using a computer-aided diagnostic system, identifying various illnesses from MRI images for effective therapy can be quite challenging. In this paper, we proposed a feature fusion of handcrafted and deep learning features for detecting and classifying brain tumors using T1-weighted MR image. In the first step, we use a genetic algorithm to select the best handcrafted features from local binary patterns (LBP) and histograms of gradients (HOG); in the second step, we use a fine-tuned conventional neural network (CNN) to identify the features of the fully connected FC layers. In the third step, we employ the feature fusion techniques of discriminant correlation analysis (DCA) and canonical correlation analysis (CCA). Extensive trials on the Radiopaedia dataset validated the classification performance. The Support Vector Machine (SVM) along with a linear kernel classifier achieved a mean accuracy of 73.88% from DCA and 93.33% from CCA on the Radiopaedia dataset. The proposed method outperforms previous recent studies in its field.