Ensemble Reads Brain Scans With 98.5% Accuracy and Shows Its Work

A study from researchers at Manipal Institute of Technology and NMAM Institute of Technology in India, published in Discover Artificial Intelligence, reports a unified deep-learning system that classifies and segments brain tumors while explaining its decisions. Led by Rashmitha R. Nayak and B. N. Anoop, the pipeline combines an ensemble classifier (soft voting across DenseNet121 and InceptionV4) for glioma, meningioma, pituitary tumor, or healthy tissue; a 3D U-Net to delineate tumor boundaries; and Grad-CAM to highlight image regions driving the output. Using a held-out test set from a Kaggle MRI dataset of 7,023 grayscale images, the ensemble reached 98.52% accuracy, outperforming InceptionV4 (98.28%), DenseNet121 (97.93%), and a lightweight baseline (95.00%).