End-to-End Brain MRI Tumor Analysis: Preprocessing, Segmentation and Classification

The goal of this work is to provide a simple and reproducible method that performs three tasks for brain MRI in one flow: consistent preprocessing, accurate tumor segmentation, and reliable multi-class classification. The procedure uses an open dataset organized for both tasks. Preprocessing converts every slice to a common size, applies luminance equalization to improve local contrast, binarizes masks with a fixed threshold, and normalizes images so the model receives stable inputs. A shared residual encoder feeds two light heads: a U-Net style decoder for pixel masks and a small classifier for four labels (glioma, meningioma, pituitary, no tumor). Training alternates segmentation and classification mini-batches with balanced losses, and model selection uses a composite of validation Dice and validation accuracy. The protocol reports all outputs needed for audit and reuse: learning curves, confusion matrices, segmentation overlays, and a history file with metrics. In a ten-epoch run, the method achieved Dice 0.800 and IoU about 0.667 for segmentation, and accuracy 0.929 with macro-F1 0.920 for classification on the test split. The approach is compact, easy to train on a standard workstation, and avoids heavy architecture. It is suitable as a dependable baseline for studies that need both a tumor mask and a case label, and it can be extended to new sites or tasks by adjusting the preprocessing or the loss balance without changing the core design.