image: a Overview of the Multi-Scale Attention Denoising Network (MSAD-Net) to recover images of high SNR from those of low SNR. b-d Typical noisy images and corresponding denoised images of (b) vascular endothelial cells (ECs), (c) MuSCs, and (d) macrophages (Mφs) (scale bar: 100 μm). Top-left panels showed the noisy images. Bottom-left panels showed the images denoised by the MSAD-Net. Top-right panels showed insets ‘(a)’ and ‘(b)’ in noisy and denoised images (scale bar: 20 μm). Bottom-right panels showed intensity along dotted lines in noisy and denoised images in top-right panels. e The original noisy image and corresponding denoised output images of MuSCs by various methods, including Gaussian filter, Median filter, BM3D, Pix2Pix, DNCNN, Masked Denoising, UNet and MSAD-Net. The image with high SNR was set as ground truth for comparison. Imaging depth: 150 μm. Imaging condition: 6 mW excitation power and 3 μs/pixel scanning time for noisy imaging while 15 mW excitation power and 12 μs/pixel scanning time (common conditions) for ground truth imaging. Scale bar: 100 μm. f Intensity profiles along the white dashed lines across MuSCs in (e). Insert was the enlarged intensity profiles in rectangular box. Three deep learning networks (Masked Denoising, UNet and MSAD-Net) with relatively good denoising effects were selected for comparison.
Credit: Yifei Li et al., Zhejiang University, PhotoniX 2025
Skeletal muscle’s ability to repair after injury or disease depends on coordinated interactions between muscle stem cells (MuSCs), macrophages, and blood vessels. However, studying these dynamics in vivo has long been limited by imaging technology: two-photon microscopy (2PM) fails at deep tissue resolution, while traditional three-photon microscopy (3PM)—though capable of 1200μm-depth imaging—requires high laser power (4–6 mW) and long scanning times (12 μs/pixel), causing photothermal damage and cell deformation.
To solve this, researchers led by Prof. Jun Qian (College of Optical Science and Engineering) at Zhejiang University developed a deep learning-enhanced 3PM system, integrating the Multi-Scale Attention Denoising Network (MSAD-Net).
Key Innovations of MSAD-Net-Enhanced 3PM
The MSAD-Net addresses low signal-to-noise ratio (SNR) issues from reduced power/time by fusing Transformer modules (for global contextual information) and Convolutional Block Attention Modules (CBAM, for local detail refinement). Trained on paired high/low SNR 3PM images of mouse tibialis anterior (TA) muscle, the network achieves:
- Ultra-low photodamage: Excitation power reduced to 1.0–1.5 mW (surface) and <20 mW (max depth)—well below damage thresholds (2.0 mW for 1300nm, 1.5 mW for 1550nm).
- Rapid imaging: Scanning time shortened to 2–3 μs/pixel, 6x faster than conventional 3PM.
- High fidelity: Average structural similarity index (SSIM) of 0.99 vs. high-power 3PM images, with 80ms/frame inference for applications.
Compared to Gaussian filtering, BM3D, or other deep learning methods (e.g., UNet, DNCNN), MSAD-Net outperforms in preserving biological structures—recovering 57 more MuSCs and 65 more macrophages in segmentation, with minimal artifacts.
Uncovering Muscle Regeneration Mechanisms
Using the enhanced 3PM system, the team achieved five-channel in vivo imaging (MuSCs: tdTomato; macrophages: GFP; blood vessels: DCBT NPs; muscle fibers: SHG; fiber membranes: THG) of mouse TA muscle over 4 days post-injury. Key findings include:
- MuSC dynamics: After injury, MuSCs proliferate at angles deviating from muscle fibers and move closer to macrophages.
- Vascular regulation: Blood vessel diameter expands in early regeneration; increased permeability causes DCBT NP leakage, triggering inflammatory factor release to recruit macrophages.
- Macrophage polarization: Macrophages shift from "responsive" (spherical) to "resting" (spindle-shaped) as regeneration progresses, guiding fiber repair.
Implications for Muscle Disease Research
This technology resolves the long-standing conflict between imaging quality and tissue safety in 3PM, providing a powerful tool for studying muscle degenerative diseases. It also sets a precedent for cross-disciplinary innovation—combining photonics, AI, and biomedicine to enable deep, low-damage imaging of other high-scattering organs.
Journal
PhotoniX
Method of Research
Experimental study
Subject of Research
Animals
Article Title
Ultra-low photodamage three-photon microscopy assisted by neural network for monitoring regenerative myogenesis
Article Publication Date
9-Oct-2025
COI Statement
The authors declare no competing interests.