Resolution-Aware Evidential Fusion for Scale-Invariant Attribution in 3D Lung Nodule Detection

Post hoc attribution is the standard transparency layer for deep learning in imaging decision support, yet its two dominant families capture conflicting evidence: activation maps are spatially coherent but coarse, while gradient-based attributions are precise but unstable. Gradient-weighted activation maps also show localisation quality that scales with lesion size, so explanation reliability degrades for the smallest lesions. We address both problems with Evidential Dempster-Shafer Fusion (EDSF), which combines an activation-based spatial prior with voxel-level path-integral attributions inside the belief-function formalism. The framework is instantiated on 3D pulmonary nodule analysis with a 3D ResNet-18 trained on LUNA16, reaching 97.2% accuracy and 0.991 mean ROC-AUC under patient-level cross-validation. On a frozen cohort of 200 candidates, EDSF is compared with Grad-CAM, GradCAM++, Integrated Gradients, Shapley additive explanations, and Occlusion Sensitivity across seven metrics of faithfulness, localisation, stability, and compactness. Localisation quality scales with nodule diameter for all five baselines, whereas EDSF shows no detectable dependence. Because this is a null result, it rests on an equivalence test rather than on a non-significant p-value: the smallest margin the data support is 0.298 for EDSF against 0.738 to 0.937 for the baselines, so practical independence holds at a medium margin and not at a tighter one. An ablation on the same two source maps shows that the property follows from the non-linear combination rather than from the choice of sources. A composite index reported with a sensitivity analysis over its weight space favours Occlusion Sensitivity under most weightings and EDSF only where stability and compactness weigh heavily.