Adaptive optical parametric optimization enables high-throughput microscopy with label-free virtual staining

Deep learning-based virtual staining offers a promising label-free alternative for digital histopathology. However, most existing methods rely on projection-based imaging and are limited to thin sections (2–5 µm) for stable image mapping. In complex pathologies, such thin sections may fail to capture sparsely distributed diagnostic features. Thick sections preserve richer tissue context but introduce severe axial information overlap in projection imaging. Here, we introduce a physically interpretable optical–computational co-design framework for high-fidelity virtual staining of ~10 µm sections. Optically, spectral–aperture coupling generates complementary measurements that jointly capture depth-selective cellular details and globally consistent tissue-scale structural information. Computationally, a physics-aligned dual-branch decoder exploits these measurements through spatial-detail and morphology-consistency pathways. Ablation studies reveal the complementary roles of spectral and aperture modulation in recovering cellular-scale details and tissue-scale morphology under severe axial overlap, while comparative benchmarking demonstrates the importance of physics-aligned decoding for translating encoded measurements into faithful morphological reconstruction. Cross-tissue evaluation across five histopathological categories and blinded pathologist assessment further verify its robustness and clinical reliability. Overall, the results establish optical–computational co-design as an effective strategy for reliable virtual staining of thick tissue sections from projection measurements.