The international journalNature Communicationshas published a study titled"Bayesian deep-learning structured illumination microscopy enables reliable super-resolution imaging with uncertainty quantification"on May 30th. This research, led by Huazhong University of Science and Technology (HUST), represents a major advancement in AI-driven super-resolution microscopy. The School of Artificial Intelligence and Automation at HUST served as the primary affiliation, with doctoral candidates Liu Tao and Liu Jiahao as co-first authors, and Professor Tan Shan (HUST) and Professor Li Dong (Tsinghua University) as corresponding authors.
Figure 1:BayesDL-SIMachieves distribution-aware super-resolution imaging
Optical super-resolution imaging technology serves as a vital tool for revealing sub-diffraction-limit information in cellular life activities, with structured illumination microscopy (SIM) being a groundbreaking instrument for live-cell super-resolution imaging. In recent years, artificial intelligence techniques like deep learning have significantly enhanced SIM's imaging quality. However, the 'black-box' nature of these methods makes it challenging to evaluate the transparency and reliability of reconstructed results. Unreliable imaging artifacts may lead to erroneous biological interpretations. Balancing reliability in SIM-based super-resolution imaging has thus emerged as a critical scientific challenge in the field.
Professor Tan Shan's team developed a Bayesian deep-learning SIM reconstruction framework (BayesDL-SIM), achieving for the first time the quantification of uncertainties in SIM-based super-resolution imaging. By modeling latent super-resolution images as high-dimensional heteroscedastic probability distributions and employing stochastic gradient Langevin dynamics (SGLD) for approximate posterior inference,BayesDL-SIMquantifies multiple types of uncertainties in imaging results. It not only provides users with pixel-level confidence assessments but also automatically identifies unreliable generalizations caused by factors such as sample type mismatches or sampling rate variations
Figure 2: BayesDL-SIM identifies unreliable super-resolution results caused by model generalization errors
Full Paper:：https://www.nature.com/articles/s41467-025-60093-w
