Super-resolution microscopy goes to single-shot learning

Understanding the dynamic behavior of living cells is essential for advancing biomedical research and drug discovery. Now, the research team led by Professor Chao Zuo at Nanjing University of Science and Technology (NJUST) have developed an AI-driven super-resolution imaging method that captures fine cellular details from a single camera exposure—a major step toward faster, less invasive biological imaging. The new technique is reported in the journal PhotoniX.

Observing how organelles, proteins, and other molecular components interact within living cells is essential for unraveling the mysteries of cellular function, disease progression, and therapeutic response. Yet capturing these intricate dynamics with the necessary clarity has remained a major challenge. In recent years, super-resolution fluorescence microscopy has expanded the frontiers of biological research, enabling scientists to visualize cellular structures and activities with unprecedented detail beyond the classical diffraction limit. However, despite these breakthroughs, current super-resolution techniques still face significant hurdles. Many approaches require capturing multiple sequential images or rely on scanning methods, leading to longer imaging times, increased vulnerability to motion artifacts, and higher photon doses—resulting in phototoxicity and photobleaching. Additionally, some techniques heavily depend on post-processing algorithms, which can produce reconstruction artifacts in complex scenarios such as low signal-to-noise ratio (SNR) conditions, resulting in reduced image fidelity. These limitations have hampered the broader application of super-resolution imaging, particularly in live-cell studies where fast, gentle, and accurate observation is critical.

To address these challenges, researchers at NJUST have developed eDL-cSIM—an ensemble deep learning–enabled, single-shot composite structured illumination microscopy technique. By employing a six-beam interference strategy, super-resolution information from multiple orientations is simultaneously encoded within a single exposure, reducing both photon dose and frame count by over ninefold. A bespoke ensemble neural network, integrating Transformer modules and multi-model fusion strategies, then deciphers the composite patterns to reconstruct images with 100 nm lateral resolution from just one frame. “Our method merges the strengths of structured illumination and deep learning to decode fine spatial information directly from minimally acquired data,” said Prof. Chao Zuo. “This not only speeds up the imaging process, but also minimizes phototoxicity, making it better suited for live-cell observation.”

As a demonstration, the researchers visualized the dynamic remodeling of intracellular structures, such as mitochondrial networks undergoing fission and fusion—processes essential for maintaining cellular health and energy homeostasis. Disruptions to this dynamic balance can lead to oxidative stress and mitochondrial dysfunction, pathological features associated with a range of diseases, including neurodegenerative disorders and metabolic syndromes such as diabetes. “By recording these processes at high resolution and frame rates, eDL-cSIM could help scientists better understand cell biology in health and disease,” said Zuo.

Moreover, eDL-cSIM excels in several critical aspects over traditional methods, such as imaging speed, reconstruction quality, and environmental robustness. It consistently demonstrates strong generalization across diverse sample types and structural features, making it highly adaptable for live-cell imaging and tracking dynamic cellular processes.

Overall, eDL-cSIM represents a paradigm shift in live-cell super-resolution microscopy. By combining clever optics with deep learning, it exemplifies a new class of "intelligent" microscopes that is both faster and gentler, opening new windows for observing subtle dynamics within living cells.