Mitigating the Resolution-Field of View Trade-Off for Comprehensive Microstructural Characterization with Advanced AI Methods.
Authors
Affiliations (5)
Affiliations (5)
- DigiM Solution LLC, 500 West Cummings Park, Suite 3650, Woburn, Massachusetts, 01801, USA.
- Office of Product Quality Research, Office of Pharmaceutical Quality, Center for Drug Evaluation and Research, U.S. Food and Drug Administration, White Oak, LS Building 64, 10903 New Hampshire Ave, Silver Spring, Maryland, 20993-002, USA.
- FDA/CDER/OPQ/OPQR White Oak, LS Building 64, 10903 New Hampshire Ave, Silver Spring, Maryland, 20993-002, USA.
- Office of Product Quality Research, Office of Pharmaceutical Quality, Center for Drug Evaluation and Research, U.S. Food and Drug Administration, White Oak, LS Building 64, 10903 New Hampshire Ave, Silver Spring, Maryland, 20993-002, USA. [email protected].
- FDA/CDER/OPQ/OPQR White Oak, LS Building 64, 10903 New Hampshire Ave, Silver Spring, Maryland, 20993-002, USA. [email protected].
Abstract
Microstructural characterization of pharmaceutical drug products is essential for understanding process-microstructure-performance relationships and ensuring consistent product quality. However, quantitative characterization is limited by a trade-off between imaging resolution and field of view: high-resolution imaging captures fine structural detail but over small sample volumes, whereas lower-resolution imaging provides broader coverage while missing critical morphological features. To address this limitation, we developed and validated an integrated framework combining convolutional neural network (CNN)-based super-resolution with Generative Adversarial Network (GAN)-based microstructure synthesis. Lyophilized drug products imaged by X-ray microscopy at multiple resolutions served as the model system. We first demonstrated that imaging resolution is a governing factor in quantitative microstructural analysis: mean pore size showed a coefficient of variation of 40.2% across resolution levels, compared with only 0.47% attributable to spatial heterogeneity within the same sample. CNN-based upscaling using ESRGAN/BSRGAN recovered solid-wall structures and pore-size distributions lost after downsampling and restored effective diffusivity toward values measured in the original high-resolution data. In an independent validation using a separate formulation imaged at 5 and 10 µm per voxel, the upscaled images reproduced pore-size distributions and diffusivity profiles closer to the 5 µm reference and showed better recovery of CQA-relevant structural features than conventional bicubic interpolation, despite lower pixel-level image fidelity scores. GAN-based synthesis expanded the field of view four-fold from a small high-resolution seed region while preserving pore-size distributions and transport properties. Together, these findings demonstrate a scalable approach for improving microstructural characterization under practical imaging constraints.