Diffusion reconstruction with validity masks and measurement consistency for low-frame structured illumination microscopy

Reducing the raw-frame budget of structured illumination microscopy (SIM) lowers exposure burden but weakens orientation–phase diversity and increases sensitivity to noise and model mismatch. We present acquisition-aware physics-guided diffusion for SIM (APD-SIM), which combines a protocol-specific conditional diffusion prior with validity-masked Poisson–Gaussian refinement. Separate synthetic-data checkpoints were trained for controller-defined digital micromirror device (DMD) protocols with three-frame (DMD-3F), six-frame (DMD-6F), and nine-frame (DMD-9F) acquisition modes; no experimental frame was used for training or fine-tuning. On the same 30 held-out fields, the three-, six-, and nine-frame APD-SIM reconstructions (APD-SIM-3, APD-SIM-6, and APD-SIM-9) achieved peak signal-to-noise ratio (PSNR) values of 31.595 ± 3.801 , 41.078 ± 2.858 , and 43.491 ± 2.085 dB, respectively. Corresponding structural similarity index (SSIM) values were 0.78525 ± 0.08267 , 0.96948 ± 0.02316 , and 0.98998 ± 0.00699 , and ground-truth-referenced Fourier ring correlation (FRC) curve areas were 0.341008 ± 0.073241 , 0.395609 ± 0.064780 , and 0.407963 ± 0.057074 . DMD-6F used one-third fewer exposures than DMD-9F. Under byte-identical DMD-6F measurements, protocol-correct masking improved final PSNR by 2.942 dB and SSIM by 0.02838 ; measurement-domain refinement improved the diffusion-only estimate by 10.96 dB and 0.398 , respectively. Relative to machine-learning structured illumination microscopy retrained for six inputs (ML-SIM-6R), APD-SIM-6 showed positive pooled image-domain effects, but their direction and magnitude depended on morphology and metric. Direct measurements from two commercial fixed-cell preparations were reconstructed by the complete two-stage pipeline using frozen protocol-specific checkpoints and fixed nominal hardware forward models. The hardware results provide qualitative evidence of zero-shot simulation-to-experiment transfer rather than a calibrated quantitative benchmark.