📝 SummarXiv

Abstract

The space-bandwidth product (SBP) imposes a fundamental limitation in achieving high-resolution and large field-of-view image acquisitions simultaneously. High-NA objectives provide fine structural detail at the cost of reduced spatial coverage and slower scanning as compared to a low-NA objective, while low-NA objectives offer wide fields of view but compromised resolution. Here, we introduce LensPlus, a deep learning-based framework that enhances the SBP of quantitative phase imaging (QPI) without requiring hardware modifications. By training on paired datasets acquired with low-NA and high-NA objectives, LensPlus learns to recover high-frequency features lost in low-NA measurements, effectively bridging the resolution gap while preserving the large field of view and thereby increasing the SBP. We demonstrate that LensPlus can transform images acquired with a 10x/0.3 NA objective (40x/0.95 NA for another model) to a quality comparable to that obtained using a 40x/0.95 NA objective (100x/1.45 NA for the second model), resulting in an SBP improvement of approximately 1.87x (1.43x for the second model). Importantly, unlike adversarial models, LensPlus employs a non-generative model to minimize image hallucinations and ensure quantitative fidelity as verified through spectral analysis. Beyond QPI, LensPlus is broadly applicable to other lens-based imaging modalities, enabling wide-field, high-resolution imaging for time-lapse studies, large-area tissue mapping, and applications where high-NA oil objectives are impractical.