Projects
Microscopy: Systems & Algorithms
We focus on developing advanced biomedical imaging technologies across modalities. Our research spans new imaging physics, bottom-up system design, and the co-design of hardware and computational algorithms tailored to specific imaging scenarios. We aim to optimize spatial, temporal, and information resolution under practical constraints such as photon or signal budget, acquisition speed, imaging depth, and system complexity.
Representative developed techniques:
(1) Sparse deconvolution is a physics-informed reconstruction framework that exploits sparsity and structural continuity to enhance resolution, contrast, and SNR beyond the limits of microscopy, particularly under low-SNR conditions. Applications include Sparse-SIM and Sparse SD-SIM.
(2) SACD is a fluctuation-based super-resolution framework that achieves approximately threefold lateral and axial resolution enhancement using only 20 frames (10 min for a 2 mm × 1.4 mm FOV), without additional optical components. Sparse-SACD further enables fast 4D live-cell super-resolution imaging. See also our label-free SACD [Light: Science & Applications, 2025].
(3) RIED is a reaction-luminescence super-resolution framework for extremely photon-limited bioluminescence (BL) and electrochemiluminescence (ECL). By decoding intrinsic luminescence fluctuations, RIED enables high-contrast 2D and 3D super-resolution imaging without external optical excitation or photochemical modulation. BL-RIED supports continuous super-resolution imaging for over two days.
From imaging physics and instrumentation to computational reconstruction and system–algorithm co-design.
Machine learning: Methods & Applications
We develop machine-learning methods for the reconstruction and analysis of multidimensional biomedical and natural signals. We exploit both explicit physical and structural priors to design interpretable computational models, and data-driven priors to develop deep-learning approaches for problems that are difficult to address with conventional methods. Our work includes self-supervised learning, representation learning, Bayesian learning, and foundation models, with applications ranging from reducing the photon, acquisition-time, and hardware requirements of microscopy to accelerating biological image reconstruction, profiling, and discovery.
Representative developed techniques:
(1) SN2N (Self-inspired Noise2Noise) is a self-supervised learning-to-denoise framework that achieves performance competitive with supervised learning while eliminating the need for large paired datasets and clean ground-truth images. It enables robust denoising across diverse biomedical imaging modalities and supports photon-efficient, long-term, and low-phototoxicity imaging.
(2) aSN2N (adaptive SN2N) introduces adaptive preprocessing and normalization strategies to stabilize self-supervised learning across heterogeneous imaging conditions. It improves robustness to intensity variations and background contamination while effectively suppressing reconstruction artifacts.
From physical and structural priors to self-supervised, representation, and foundation-model learning.
Smart Bioimage Analysis & Biomedical Devices
We develop intelligent bioimage analysis methods and biomedical imaging instrumentation to enable automated, quantitative, and high-throughput biological investigation. Our work spans image segmentation, tracking, profiling, and representation learning, as well as the development of compact, application-oriented imaging devices and systems.
Representative developed techniques:
(1) PANEL (pixel-level analysis of error locations) is a quantitative framework for mapping reconstruction errors at the super-resolution scale, enabling systematic assessment of both conventional and deep-learning-based reconstruction methods.
(2) Smart palm-size optofluidic hematology analyzer is a compact imaging-based platform for automated leukocyte concentration measurement and rapid hematological analysis.
(3) Light-field endoscopic probes initiated with MIRD (multiple micro-imaging devices), provide highly flexible and compact probes for single-shot, high-resolution volumetric imaging. We continue to develop new optical designs and computational imaging strategies to improve imaging performance, probe flexibility, and practical applicability. See also [Optics and Lasers in Engineering, 2026] and [Optics Express, 2026].
(4) Sparse confocal microscopy with single-particle tracking was applied to reveal how SARS-CoV-2 virus-like particles exploit filopodia to reach cellular entry sites through two distinct modes, “surfing” and “grabbing”, thereby reducing random searching on the plasma membrane. See also [European Journal of Cell Biology, 2025].
From quantitative bioimage analysis to compact and application-oriented biomedical devices.
