Accelerating Large-Area Ultrastructural Imaging of Bone Tissues by Integrating SEM-BSE with Diffusion-Based Deep Learning — ASN Events
  1. Schedule
  2. Christopher and Margie Nordin Young Investigator Award
  3. Accelerating Large-Area Ultrastructural Imaging of Bone Tissues by Integrating SEM-BSE with Diffusion-Based Deep Learning

Accelerating Large-Area Ultrastructural Imaging of Bone Tissues by Integrating SEM-BSE with Diffusion-Based Deep Learning (#111)

Qiongxiang Q.L. Lin 1 , Xiaojun X.C. Chen 1 , Killugudi Swaminatha K.I. Iyer 1 , Haibo H.J. Jiang 2 , Kai K.C. Chen 1
  1. The University of Western Australia, Crawley, WA, Australia
  2. The University of Hong Kong, Hong Kong

Ultrastructural imaging is fundamental to understanding bone matrix organisation and cellular function. Scanning electron microscopy with backscattered electron detection (SEM-BSE) allows high-resolution imaging over large tissue areas and is well suited for mineralised skeletal tissues. Yet imaging throughput remains constrained by the trade-off between acquisition speed and image quality: shorter pixel dwell times accelerate scanning but introduce noise that obscures subcellular detail. Here, we integrate EMDiffuse, a diffusion-based deep learning algorithm, with our established SEM-BSE workflow to decouple acquisition speed from image quality and accelerate large-area ultrastructural imaging of bone tissues.

Mice bone tissues were processed using our previously established heavy metal staining and resin embedding protocol. Ultrathin (500 nm) sections on silicon wafers were imaged using an FEI Verios SEM with a BSE detector, and MAPS software automated tile acquisition. To train the EMDiffuse denoising model (EMDiffuse-n), we acquired paired images of the same regions at short and standard dwell times. The pre-trained model was adapted to bone tissue by fine-tuning on a single pair of training images.

We applied this approach to trabecular bone, cortical bone, and bone marrow. EMDiffuse-n recovered ultrastructural details, including osteocyte lacunae, collagen fibril organisation, and subcellular organelles, from rapidly acquired noisy images. Output quality was comparable to conventional long-dwell-time acquisitions. The model also generates pixel-wise uncertainty maps for each prediction, so that users can flag regions that may need closer inspection. Adaptation to bone tissue required only one fine-tuning image pair, confirming the model's transferability to musculoskeletal samples.

This workflow addresses a throughput limitation in large-area SEM-BSE imaging. Rapid acquisition without loss of ultrastructural detail makes high-throughput morphometric and spatial analysis of bone tissues feasible, with applications in skeletal development, ageing, and disease research.