Deep Feature Fusion and Ensemble Learning to Create an Effective CNN Brain Tumour Classification Model

Early brain tumour exploration can streamline treatment. Some automated diagnosis system aids radiologists in distinguishing between normal and abnormal brain tissues, simplifying clinical and diagnostic processes. However, categorising MRI images is challenging due to low contrast, noise, tumour shape and localisation dissimilarity, and similarity between ordinary and cancerous regions of interest (ROIs). This study uses a deep convolutional neural network with feature blending and ensemble learning to analyse MRI abnormalities, followed by detection and classification tasks. The ensemble learning method effectively distinguishes between ordinary and cancerous tumour ROIs, yielding reliable results. Feature fusion identifies discriminative features between classes. To address overfitting in smaller datasets, depth-wise separable convolution and spatial drop-out techniques are explored for MRI brain image classification. The proposed approach has been validated on two freely available datasets, Kaggle and BrATS, with the BrATS dataset showing superior outcomes in accuracy, specificity, and sensitivity (0.995, 0.996, 0.996).