Averaging-Based Hybrid Ensemble of DenseNet-121 and ResNet-50 for Computer-Aided Brain Tumour Diagnosis: A Step Towards Early Detection

Abstract

One of the deadliest diseases that can affect children and adults is a brain tumour. A brain tumour is the abnormal growth of cells in the brain or near it. Commonly, there are two types of tumours: malignant (cancerous) and benign (non-cancerous). The importance of early detection of brain tumours cannot be overstated, as tumours progress rapidly and decrease survival rates. Due to the complexity of brain tumour structure and classification, manual inspection can lead to false-positive or false-negative results. The complexity of tumour detection can be reduced by using deep learning segmentation and classification models. This paper proposes an AI-driven solution for screening brain tumours using MRI images, emphasising improving the model’s sensitivity to reduce missed tumour cases. A deep hybrid ensemble model is developed by combining two CNN models using an averaging ensemble learning technique. The models were trained and evaluated on a publicly available MRI dataset consisting of 4,845 images. Four state-of-the-art deep learning architectures were trained and evaluated, and the best-performing models were combined through the ensemble technique. Our final model achieves 98.49% accuracy and 97.58% sensitivity. Compared with a recent ensemble-based baseline, the proposed approach reduces error rate and false negative cases, improving robustness and clinical reliability.