Leakage‐Controlled Evaluation of Pixel‐Level MRI–sCT Fusion Reveals a Quality–Performance Mismatch in Brain Tumour Classification

In medical image classification, pixel‐level fusion has been presupposed to be useful in merging complementary visual information and improving classification performance. In this research, we test this assumption by considering fused‐image quality and image‐level leakage‐controlled classification performance for MRI–sCT brain tumour analysis. Seven EfficientNet‐B0 input representations were evaluated on 1,469 test image pairs from four classes (glioma, meningioma, no‐tumour and pituitary). Because patient identifiers were unavailable, patient‐level independence could not be verified. To assess the stability of single‐run performance estimates, all representations were retrained using three random seeds, and MRI‐only comparisons were assessed using 1,000‐iteration bootstrap resampling, McNemar's paired testing and Holm–Bonferroni correction. Across seeds, AVG50 fusion achieved the highest mean accuracy and macro‐F1, while MRI‐only remained statistically comparable. None of the pixel‐level fusion methods was significantly superior to MRI‐only after correction, whereas sCT‐only was consistently weaker. The quality–performance analysis showed that fused images with visually or statistically richer appearance did not necessarily produce better classification performance. MRI‐fidelity measures, particularly SSIM and edge preservation versus MRI, were more closely associated with performance, whereas mutual information versus MRI showed a moderate‐to‐strong but non‐significant association; entropy and sharpness were not significant predictors. The MRI‐only model showed anatomically plausible attention in Grad‐CAM analysis of representative cases. These results suggest that pixel‐level MRI–sCT fusion should not be considered superior to MRI‐only classification in this dataset, but that conservative fusion may still be useful when an enhanced single‐image MRI–sCT representation is needed without expectation of significant diagnostic‐performance gain.