Adaptive saliency cascade network for efficient multiclass classification of multi-source brain MRI data

Multi-source brain MRI classification is complicated by subtle inter-class similarities and source-dependent variation in contrast, orientation, field of view, and acquisition protocol. These factors can obscure disease-related cues and make apparent performance sensitive to the evaluation design. This study proposes the Adaptive Saliency Cascade Network (ASC-Net), a compact four-stage classifier built from Adaptive Saliency Conduit (ASC) blocks that combine efficient additive attention with gated feature modulation for selective spatial and channel refinement. ASC-Net achieved 99.34% accuracy and a 99.36% macro F1-score on the randomly partitioned ND-5 image-level benchmark, exceeding the strongest directly trained comparator by 0.71 and 0.69 percentage points, respectively, with 11.15 million parameters and 4.19 GFLOPs. Because patient or examination identifiers are unavailable for the consolidated ND-5 sources, this internal result is interpreted strictly as an image-level benchmark. Generalization was examined separately on BDNeuro-MRI, an external dataset-level cohort excluded from training, validation, checkpoint selection, and hyperparameter decisions. Without retraining or adaptation, ASC-Net obtained 98.70% accuracy (95% CI: 98.38%–98.96%) and a 98.61% macro F1-score (95% CI: 98.28%–98.92%) on 5941 external images. In post-hoc exploratory image-level comparisons, the macro F1 difference from the strongest external comparator remained significant after Holm correction ( 𝑝 H o l m = 0 . 0 0 0 4 ). Patient- or case-level identifiers were unavailable for clustered external resampling, so the reported confidence intervals and 𝑝 values quantify image-level rather than patient-level uncertainty. Ablation analysis preserved the same component ordering on ND-5 and BDNeuro-MRI. These findings support cross-dataset transfer while retaining the need for verified patient-level and prospective clinical validation.