Adaptive saliency-guided signal representation for robust brain tumor MRI classification

Magnetic Resonance Imaging (MRI) analysis is often affected by acquisition noise, intensity inhomogeneity, and complex anatomical variations that complicate the extraction of discriminative signal patterns. Although deep learning methods have demonstrated strong performance in medical image analysis, many approaches treat images as visual patterns without explicitly modeling their underlying signal structure. This paper proposes an Adaptive Saliency-Guided Signal Representation (ASGSR) framework that explicitly models MRI data as structured spatial signals. The proposed framework formulates MRI classification as a task-conditioned signal-representation problem in which classification-derived spatial sensitivity determines the subsequent spatial transformation of the MRI signal. The resulting saliency-conditioned representation is subsequently characterized using multi-resolution and statistical operations before probabilistic classification. A Bayesian decision model is employed for classification with posterior-probability-based confidence estimation. Experimental evaluation on multiple publicly available brain MRI datasets demonstrates improved classification performance and robustness under the evaluated conditions compared with the selected convolutional and transformer-based baselines. The method achieves an average classification accuracy of 95.2% and an AUC of 0.97 across the evaluated datasets, while the derived spatial representation exhibits higher PSNR and SSIM under the defined representation-level evaluation protocol. These results indicate that the proposed signal-centric formulation can improve classification performance, robustness to the evaluated perturbations, and the spatial sensitivity representation of MRI inputs under the evaluated experimental conditions.