The electron microscope (EM) has transformed our capacity to observe the detailed structures within cells. The evolution to three-dimensional electron microscopy, termed volume EM (vEM), has significantly improved our ability to capture nanoscale images in three dimensions. Nonetheless, challenges related to imaging speed, quality, and sample size continue to restrict the spatial area and volume that can be effectively imaged. Meanwhile, artificial intelligence (AI) is increasingly becoming a crucial player in various scientific fields, facilitating discoveries and acting as an essential resource in research methodologies.
In light of recent advancements in AI-driven image generation technologies—particularly through sophisticated diffusion models—a research team led by Professor Haibo JIANG from the Department of Chemistry and Professor Xiaojuan QI from the Department of Electrical and Electronic Engineering at The University of Hong Kong (HKU) has created a set of algorithms named EMDiffuse. This novel approach seeks to boost imaging capabilities and tackle the trade-offs encountered in EM and vEM. Their research has been published in Nature Communications.
For standard 2D EM, EMDiffuse significantly improves the restoration of authentic, high-quality images, showcasing detailed ultrastructural aspects even when starting from images with noise or low resolution. Unlike other deep learning techniques focused on denoising or super-resolution, EMDiffuse employs a distinctive methodology where the solution is sampled from the desired distribution. In this way, low-quality images are utilized as a guiding factor at each stage of the diffusion process, thereby ensuring that the generated structure remains accurate. This means the low-quality input actively shapes the restoration rather than merely serving as a starting reference. The diffusion model effectively combats blurring, achieving a level of resolution that approaches true representations, which is vital for detailed ultrastructural analyses. Additionally, EMDiffuse’s versatility allows it to be applied directly to different datasets or with minimal fine-tuning using just a single pair of training images.
In the realm of vEM, current technology often finds it challenging to capture high-resolution 3D images of larger specimens, particularly in depth (‘z-direction’), hindering comprehensive studies of critical 3D cellular structures such as mitochondria and the endoplasmic reticulum.
EMDiffuse offers two adaptable strategies to overcome this challenge. It can utilize ‘isotropic’ training data—3D datasets with consistent, high resolution across all dimensions—to learn and enhance the axial resolution of other 3D data. Alternatively, EMDiffuse can evaluate and improve the depth resolution of existing 3D images using self-supervised methods, which do not require specialized training data. This flexibility enables EMDiffuse to significantly elevate the quality and applicability of 3D electron microscopy data for various research purposes.
The enhanced volumes reveal remarkable precision in examining ultrastructural features, such as mitochondrial cristae and the interactions between mitochondria and the endoplasmic reticulum, which can be difficult to observe in the original anisotropic volumes. Because EMDiffuse does not rely on isotropic training data, it can be directly applied to any available anisotropic volume to enhance its axial resolution.
EMDiffuse marks an essential leap forward in the imaging potential of both EM and vEM, augmenting the image quality and axial resolution of the captured data. ‘Building upon this foundation, we anticipate further advancements and rapid progression of the EMDiffuse algorithm, opening the door for comprehensive explorations into the complex subcellular nanoscale ultrastructures within large biological entities,’ stated Professor Haibo Jiang, one of the paper’s corresponding authors. ‘As this AI-driven imaging technology evolves, we are eager to witness its role in helping researchers discover previously unexplored operational mechanisms in biological systems,’ added Professor Xiaojuan Qi, also a corresponding author of the paper.
