Spatial Statistics and Explainable Deep Learning for 3D Cell–Protein Interaction Profiling

Digital histology of human postmortem brain tissue is fundamental to understanding neurodegeneration, yet resolving the spatial organisation of neuroinflammatory markers remains challenging because manual scoring and scalar summaries capture only part of the image information. We present a dual-tiered computational framework combining stochastic spatial point-process analysis with explainable deep learning. Using high-resolution spinning-disk confocal immunofluorescence images from a human Parkinson’s disease cohort (Braak Stage 3/4), we analysed spatial associations between IBA1-positive microglia and phosphorylated alpha-synuclein (pSyn) aggregates across the nigrostriatal pathway. Second-order spatial point-process statistics did not detect robust group differences in microglial spatial clustering, while a voxel-level Overlap Index showed a nominal elevation of microglial–pSyn co-occupancy in the substantia nigra that did not survive multiple-testing correction (d=0.74, nominal p=0.045, BH-adjusted p=0.550; diagnosis-by-region interaction p=0.464). To test whether pixel-level image information provided donor-level discrimination, we trained 2.5D multi-slice ResNet-18 models under repeated donor-contained cross-validation across all five outer folds, with predictions aggregated to the donor level. In the substantia nigra, the pSyn-only model provided the most consistent discrimination (mean AUROC 0.694±0.062 across seeds; seed-averaged donor out-of-fold AUROC 0.669, 95% CI 0.405–0.901), whereas adding IBA1 did not yield measurable incremental discrimination (ΔAUROC −0.074, 95% CI −0.248 to 0.083; p=0.395). Putamen performance was near chance and unstable across seeds. Cohort-level quantitative attribution analysis indicated that relevance maps were associated with structures present in the input images, but no disease-specific attribution differences survived multiple-testing correction. The framework provides a donor-aware protocol for comparing interpretable spatial summaries, imaging modalities, and held-out explanations; external validation in larger cohorts is required before generalisation.