Trust: A Unified Attention‐Based and Uncertainty‐Aware Framework for Brain Tumor Detection and Grading

Abstract Artificial intelligence (AI) system design for brain tumor diagnosis by utilizing Magnetic Resonance Imaging (MRI) demands more than accuracy; it also entails the development of robust, understandable, and medically trustworthy systems capable of handling heterogeneous imaging situations. The existing models are usually task‐oriented, that is, either segmenting or classifying brain tumors in MRI images, and cannot easily generalize when faced with images from other machines and procedures. With the goal of tackling these limitations, this research seeks to develop the TRUST Brain Intelligence Framework. This proposed model is a unified architecture integrating Adaptive Contrast Harmonizer, attention‐enhanced DeepLabV3+ segmentation, EfficientNetV2‐based Scale Fusion Grader, Monte Carlo Dropout uncertainty estimation, and Grad‐CAM++ explainability. Entropy‐driven pre‐processing enhances intensity consistency and tumor boundary preservation, whereas attention‐guided segmentation and adaptive multi‐scale grading facilitate discriminative feature learning. Empirical results using the BRISC and Brain Tumor MRI data sets reveal that the proposed approach significantly outperforms traditional approaches based on convolutional neural networks and segmentation. The proposed model was found to be highly effective in terms of Dice, Intersection Over Union, accuracy, and F1‐score, and it exhibited consistent performance across heterogeneous MRI modalities. Uncertainty‐aware predictions and visual explanations further improved clinical interpretation of the results, making the framework more suitable for trustworthy AI‐assisted neuro‐oncology decision support.