MULTI-SCALE FEATURE FUSION AND TRANSFORMER-BASED COMPUTER VISION FOR AUTOMATED GLIOMA GRADING AND BRAIN TUMOR SEGMENTATION IN MRI IMAGES

Abstract

Accurate glioma grading and precise brain tumor segmentation are essential for early diagnosis, effective treatment planning, and improved clinical management of patients with brain tumors. However, the heterogeneous morphology, irregular tumor boundaries, overlapping tissue characteristics, and complex spatial variations observed in magnetic resonance imaging (MRI) present significant challenges for conventional computer-aided diagnostic systems. This study presents an advanced multi-scale feature fusion and Transformer-based computer vision framework designed to achieve automated glioma grading and accurate brain tumor segmentation through multimodal MRI analysis. The investigation utilizes a dataset comprising 8,400 MRI examinations from 2,100 patients, including 1,260 high-grade and 840 low-grade glioma cases, incorporating T1-weighted, contrast-enhanced T1, T2-weighted, and FLAIR imaging modalities. The dataset was partitioned at the patient level into training (70%), validation (15%), and testing (15%) subsets to ensure reliable performance assessment. A comprehensive preprocessing pipeline involving image registration, skull stripping, intensity normalization, and data augmentation was implemented to improve image quality and standardize anatomical representations. The proposed architecture combines convolutional neural networks (CNNs) for hierarchical feature extraction, multi-scale feature fusion for preserving fine-grained tumor characteristics, and Vision Transformer (ViT) mechanisms for capturing long-range spatial dependencies. A unified multi-task learning strategy enables simultaneous tumor-region segmentation and glioma-grade classification. The reported experimental results demonstrate 97.42% classification accuracy, 97.16% precision, 96.88% recall, 97.02% F1-score, and an area under the receiver operating characteristic curve (AUC) of 0.989. Furthermore, the segmentation framework achieves Dice similarity coefficients of 94.86%, 92.73%, and 90.41% for whole tumor, tumor core, and enhancing tumor regions, respectively. Comparative experiments against U-Net, Attention U-Net, conventional CNN architectures, and standalone Transformer models demonstrate superior segmentation precision and grading performance. Ablation analysis further establishes the contribution of hierarchical feature integration and attention-driven contextual learning. Overall, the findings highlight the potential of the proposed framework to enhance diagnostic consistency, support automated neuroimaging interpretation, and advance intelligent clinical decision-support systems for personalized glioma management.