NeuroTürkNet: A Two-Stage Deep Learning Framework for Brain MRI Tumor Classification and Localization

Brain magnetic resonance imaging (MRI) presents challenges for computational tumor classification and localization because tumor appearance and surrounding anatomy vary across examinations. This work presents NeuroTürkNet, a two-stage prototype that combines a custom convolutional neural network (CNN) for four-category image classification with a YOLOv8n model for tumor-region localization. Input MRI images were resized to 640 × 640 pixels and processed using contrast-limited adaptive histogram equalization (CLAHE) and bilateral filtering. The classification dataset comprised 3,264 publicly sourced images, and the localization dataset comprised 5,858 annotated images. The CNN was trained with Adam optimization, class weighting, and validation-based learning-rate scheduling; the YOLOv8n component was trained separately and applied to images classified as tumor-bearing. In the detailed test results reported for the classifier, accuracy was 90.84%, precision 90.87%, recall 90.84%, F1-score 90.80%, and ROC-AUC approximately 0.99. The localization component reported an Intersection over Union (IoU) of 0.81, a best mAP@50 of 0.88, and best precision of 0.85. A graphical prototype supports MRI upload, visualization of predicted regions, and PDF output. The results demonstrate the feasibility of the proposed computational workflow on the datasets described, but they do not establish clinical diagnostic validity. Independent validation and a complete audit of dataset partitions are needed before medical deployment.