Trustworthy Explainable AI for Cross-Dataset Brain Tumor Diagnosis from MRI

Diagnosing brain tumors through magnetic resonance imaging (MRI) is of great importance, correctly and reliably because of the domain shift, uncertainty in the models, and even because of insufficient interpretation of automated systems. In this paper, an integrated credible and explicable framework of artificial intelligence aimed at incorporating tumor-sensitive representation learning, exposition-steadfast feature selection, uncertainty-senient classification, and selective prediction coupled with probability modification, to aid sound clinical decision representation, is presented. On a cross-dataset evaluation protocol, which trains the models using the Figshare Brain Tumor Dataset and test them using the BraTS MRI Dataset, the proposed framework shows an accuracy of 92.37% and area under the ROC curve (AUC) of 0.948, with substantially better calibration reliability, with an expected calibration error of 0.031, in comparison to hybrid, convolutional, and transformer baselines. Quantitative explainability tests show more faithfulness (0.74) and stability (0.81) scores compared to other competing methods. Reliability maps and visualization illustrations also show consistent estimates of confidence and attention of the models focused on the lesions. These findings support the fact that the framework is always improving the predictive performance, calibration reliability, and the explanation quality across architectures. The future research will broaden the assessment to bigger multi-centric studies on clinical cohorts, study semi-supervised pretraining techniques, in addition to causal interpretability using human-in-the-loop mechanisms to enhance further the bridge between automated reasoning and expert clinical delineations.