Medical deep neural networks (DNNs) trained upon coarse image inputs are inherently insensible to fine-grained anatomic features. To enhance DNN perception on delicate microvascular structures, we proposed using a straightforward angiographic mobile-dense hybrid network (AMDenseNet) in tandem with a flexible input split, suppression, and swap perturbation (InSSS-P) framework to perform explainable diagnostics for microvascular diseases. Mechanistically, InSSS-P and AMDenseNet conjointly (1) decompose complex anatomy inputs into LEGO-like blocks, then (2) distill plexus-wise vascular block significance from the rearranged anatomy input-output samplings. To validate the robustness of the proposed model, we trained AMDenseNet to detect blind-threatening retina exudative age-related macular degeneration (exAMD) from the micrometer-scaled optic coherence tomographic angiography (OCTA), wherein our AMDenseNet annotated triple-specific (plexus, branch, and exudate activity) neovascular (NV) features at a diagnostic precision that is comparable to the gold standard dye-based angiography. Interestingly, when we applied plexus-wise InSSS-P to the AMDenseNet predictions, the permutate results demonstrated that input perturbation at deep capillary plexus (DCP) layer abrogated model-awared exAMD features and resulted in false-negative exAMD impressions. On the contrary, choroid capillary (CC) input perturbations contributed to false-positive exAMD characterization. Notably, the AMDenseNet triple-specific NV annotation functionality was disrupted by either DCP or CC input perturbations, indicating the significance of DCP and CC in the task of exAMD characterization. In sum, this study employed anatomy decomposition approaches to distill plexus-wise microvascular significance and highlighted the analytic potential of leveraging an anatomy-sensible model to discover novel disease biomarkers.