Lightweight Uncertainty‐Calibrated Imaging by Deep Learning for Reliable and Rapid One‐Shot Super‐Resolution Microscopy

ABSTRACT Deep learning (DL) has significantly accelerated Fourier ptychographic microscopy (FPM) reconstruction, yet existing data‐driven approaches universally produce deterministic outputs without any measure of prediction confidence — a critical barrier to the clinical and quantitative adoption of DL‐based FPM. Here we present Lightweight Uncertainty‐Calibrated Imaging by Deep learning (LUCID), which is a lightweight framework that jointly achieves high‐resolution FPM reconstruction from a single raw input image and performs pixel‐wise uncertainty quantification. Compared with traditional iterative methods, the model achieves high‐fidelity reconstruction with only 0.71 million (M) parameters and delivers a speedup of more than 100×. In experiments on the FPM‐BioCell dataset, which spans 10 biological specimen categories, LUCID outperforms the state‐of‐the‐art model WM‐FPM (despite being over two orders of magnitude smaller) and other competing models in terms of PSNR, SSIM, and MS‐SSIM, while, consistent with the perception–distortion trade‐off, the adversarially trained WM‐FPM attains a better perceptual LPIPS. Crucially, the predicted uncertainty maps are well‐calibrated. Thus, they provide spatially resolved confidence indicators that can be used to identify which regions of a reconstruction are trustworthy and which require cautious interpretation. Further experiments on the Bio‐SR dataset validate its applicability to broader super‐resolution imaging tasks.