Contamination-Aware Cross-Source and Cross-Architecture Benchmarking of Explainable AI for Brain-Tumor MRI Classification

Brain-tumor MRI classification and explainable artificial intelligence (XAI) evaluation can be affected by hidden source overlap and architecture-dependent explanation rankings. This study presents a focused contamination-aware, cross-source, and cross-architecture sensitivity study for brain-tumor MRI classification and post-hoc XAI. Three ImageNet-pretrained classifiers—ResNet-18, DenseNet-121, and ConvNeXt-Tiny—were trained exclusively on 3064 Cheng contrast-enhanced T1-weighted MRI slices using three random seeds. External evaluation used 4793 BRISC tumor images. A preceding image-level audit flagged 1070 BRISC images as possible or probable Cheng-related overlaps; therefore, performance was reported on BRISC-Full, an audit-retained subset of 3723 images, and the 1070-image flagged cohort. DenseNet-121 provided the strongest and most reproducible cross-source classification, achieving 90.34 ± 1.77% accuracy, 90.25 ± 1.79% macro-F1, and 0.9927 ± 0.0009 macro ROC-AUC on BRISC-Conservative. Removing flagged images reduced macro-F1 by 1.05–1.88 percentage points across architectures, while the flagged cohort was systematically easier, showing sensitivity to cohort selection without establishing a causal leakage effect. A comprehensive 15-method XAI benchmark on a pre-specified ResNet-18 showed that no technique dominated faithfulness, stability, and efficiency simultaneously: HiResCAM and LayerCAM offered the strongest overall quality–efficiency balance, whereas ScoreCAM achieved excellent stability at markedly higher computational cost. An eight-method common-correct cross-model benchmark revealed significant rank agreement between ResNet-18 and DenseNet-121 (Spearman ρ = 0.731, p = 0.040) and weaker estimated agreement involving ConvNeXt-Tiny, where Integrated Gradients became the leading method. Additional five-fold patient-separated Cheng evaluation yielded mean macro-F1 values of 0.913–0.936. In a distinct 12-method panel of 300 held-out Cheng images, overall rankings on 298 complete images gave ResNet–DenseNet agreement of 0.965 (95% CI 0.867–0.970), compared with 0.098 and 0.147 for pairs involving ConvNeXt. XAI features did not consistently improve a patient-separated error-referral task. These results support architecture- and protocol-specific interpretation of explanation reliability within the tested datasets.