Brain Tumor Detection and Localization Using Advanced Deep Learning Ensembles and YOLO-Based MRI Analysis

Abstract

The detection of brain tumors is a vital endeavor in medical imaging, since the precise classification and localization of tumor areas in MRI images significantly affect diagnostic accuracy and treatment strategies. This study employs the Brain Tumor MRI dataset for multi-class classification and detection tasks. Advanced transfer learning architectures such as EfficientNetV2, Vision Transformer (ViT-B16), DenseNet121, and Xception were utilized for classification, supplemented by ensemble techniques based on average, weighted average, and geometric mean to improve predictive reliability. Multiple YOLO versions (YOLOv5, YOLOv8, YOLOv9, and YOLOv11) were employed for tumor localization and identification, utilizing bounding box annotations in YOLO format. Furthermore, explainable AI methodologies, like Grad-CAM, were utilized to produce heatmaps that emphasize the discriminative tumor locations affecting model predictions. Experimental results indicate that the DenseNet121 and Xception models attained the greatest classification accuracy of 99.54%, surpassing other classifiers. YOLOv8 exhibited exceptional performance in detection, with a mean average precision (mAP) of 95.1%. These findings confirm the efficacy of combining classification and detection frameworks with interpretability methods for accurate and dependable brain tumor diagnosis. A Flask-based web application was created to deploy the trained models, allowing users to upload MRI images and examine tumor localization, predicted classifications, and confidence scores via an interactive interface