Multi-Scale Neuro Diffusion Transformer Framework with Graph-Attentive Patch Learning For Robust Brain Tumor Classification In Mri

Abstract

Brain tumor classification from the MRI scans is remaining as a challenging task because of high inter-patient variability, irregular tumor morphology, and inconsistent imaging quality across the datasets. conventional deep learning models often depend on large-scale labeled datasets and struggle in maintaining robustness when confronted with the noise, modality shifts, and class imbalance. In clinical settings, this limitation is reducing the reliability and is restricting deployment in real diagnostic workflows. Despite recent advances in deep neural architectures, most of the Conventional approaches is failing to jointly address three critical issues: degraded MRI quality because of acquisition noise, limited annotated medical data, and insufficient spatial-context modeling of tumor regions. Further, the current systems often treat feature extraction and classification as separate stages, which is introducing redundancy and is reducing the interpretability under the heterogeneous imaging conditions. This study is introducing a Multi-Scale NeuroDiffusion Transformer Framework with the Graph-Attentive Patch Learning (MND-GAPL). The framework first is applying an adaptive frequency-guided restoration module that is improving the structural consistency while it is suppressing the imaging noise without relying on the handcrafted filters. To mitigate dataset imbalance, a diffusion-driven augmentation mechanism generates anatomically consistent synthetic MRI samples, preserving tumor boundary realism. The core of the framework is combining the patch-based transformer encoder with the graph-attentive relational learning, where each image patch is modeled as a node in a spatial dependency graph. This design is allowing dynamic interaction between the local tumor textures and global anatomical structures. A self-supervised contrastive objectiverefines feature separability by aligning latent representations across the augmented views, which is improving generalization under the limited supervision. Finally, an uncertainty-aware decision head is providing calibrated predictions that is considered suitable for clinical interpretation. The proposed MND-GAPL framework is evaluated on Figshare, Kaggle Brain MRI, and T1-weighted contrast-enhanced datasets. It is achieving an accuracy of 98.3%, sensitivity of 97.9%, specificity of 98.6%, F1-score of 98.1%, and ROC-AUC of 0.997, which is performing better than ResNet50V2, EfficientNetB3, and Swin Transformer. Cross validationis confirming stability with the 98.3% ± 0.2 mean accuracy, while statistical testing is showing significant improvement with the p < 0.001. These results is showing superiorconvergence stability, and clinical reliability under the heterogeneous MRI conditions.