BrainNet-S: A Fully Automated Hybrid Stacking Classifier for Enhanced Brain Tumor Classification and Prediction

Brain tumors are abnormal masses of tissues that grow uncontrollably. Early identification warrants treatment plans to fight the tumors before they grow too large. MRI scans are a common approach for identifying them. However, it’s complicated for healthcare experts to identify, segment, and remove the tissue regions, making finding the tumors labor-intensive and inefficient. This elicits machine learning techniques to streamline this process and aid in early diagnosis. Machine learning techniques have many applications for challenging medical issues like brain tumor detection. This research proposes a novel, fully automated stacking classifier that uses a CNN (DenseNet121) as a feature extractor. It takes care of all data augmentation on runtime, preparing it for real-world scenarios. The features extracted by the CNN are parsed into an XGBoost and SVM classifiers, independently tuned using a random search. Finally, their predictions are combined in a logistic regression meta-classifier for multiclass, which produces the final output. BrainNet-S aims to comprehensively leverage the individual strengths of each model to create a more robust and efficient classifier. The model was trained on the figshare brain tumor repository containing 3064 images of brain MRI scans with and without tumors [12]. The model performed exceptionally well and received a specificity of 0.992, recall of 0.988, F1-score of 0.986, and precision of 0.985. It also received an overall accuracy of 0.9837 on the testing set. These findings highlight BrainNet-S as a promising alternative to traditional medical and ML techniques, offering enhanced performance and efficiency. Employing this algorithm could benefit hundreds of thousands through early diagnosis.