FALCON: A multi-layered fusion ensemble for high-certainty brain tumor diagnosis from MRI

Accurate classification of brain tumors from Magnetic Resonance Imaging (MRI) is essential for patient prognosis, yet single deep learning models often lack the robustness for clinical certainty. This study introduces FALCON (Fusion of Architecturally Layered Classifiers for Oncology), a novel framework designed to overcome these limitations through a systematic, multi-layered fusion of heterogeneous deep learning architectures. The framework was evaluated using a four-class MRI dataset (glioma, meningioma, pituitary, and no tumor) containing 7,023 images. We first established a baseline by training a diverse cohort of 11 base classifiers: three bespoke Convolutional Neural Networks (CNNs) and eight fine-tuned EfficientNet variants (B0-B7). While the top standalone models achieved excellent accuracies (up to 99.69%), our multi-layered approach demonstrated clear performance gains. Homogeneous ensembles of similar models improved the accuracy to 99.85%. Critically, the final layer of the FALCON framework, which fuses architecturally diverse custom CNNs and EfficientNets, achieved a flawless 100% classification accuracy. These results establish that a systematic, multi-layered fusion of heterogeneous models is a superior strategy for maximizing diagnostic certainty and provides a robust methodology for developing clinical-grade AI in neuro-oncology. • Custom CNNs and EfficientNets were tested for brain tumor MRI classification. • Fine-tuned EfficientNet-B3 achieved the highest single-model accuracy (99.69%). • Ensembles of EfficientNets consistently achieved a superior 99.85% accuracy. • A novel fusion of architectures (FALCON) yielded 100% classification accuracy. • The framework offers a robust model for clinical decision support in oncology.