Advanced Photonics Nexus, Volume. 4, Issue 3, 036007(2025)

PLayer: a plug-and-play embedded neural system to boost neural organoid 3D reconstruction

Yuanzheng Ma1,†... Davit Khutsishvili1, Zihan Zang2, Wei Yue3, Zhen Guo4, Tao Feng1, Zitian Wang1, Liwei Lin3, Shaohua Ma1,* and Xun Guan1,* |Show fewer author(s)
Author Affiliations
  • 1Tsinghua University, Tsinghua Shenzhen International Graduate School, Shenzhen, China
  • 2University of California, Department of Bioengineering, Los Angeles, California, United States
  • 3University of California, Department of Mechanical Engineering, Berkeley, California, United States
  • 4Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, Cambridge, Massachusetts, United States
  • show less
    Figures & Tables(7)
    Preprocessing and image acquisition of neural organoids. The sequence follows, in a clockwise direction: neural organoid culturing and preparation, organoid cryosectioning, slide selection, primary and secondary antibody application, confocal microscopy, layer application of PLayer neural network, and 3D reconstruction. Stained with anti-beta III tubulin primary, and a secondary antibody conjugated to 488 nm. Some components of the figure are created with BioRender.com.
    Schematic of the PLayer. (a) Sequential stages of the process: (I) GAN-based denoising, (II) SME interpolation, (III) restorer, and (IV) model optimization and deployment. (b) IPF interpolation strategy incorporates SME to estimate the probability of a pixel appearing on the target layer from neighboring layers. (c) Pruning optimization flow chart: we increase the pruning ratio until improvements plateau; further pruning degrades performance. Stained with anti-beta III tubulin primary, and a secondary antibody conjugated to 488 nm. W, weight; D, distance; Acc, accuracy; Exp., expected accuracy threshold.
    Iterative resolution enhancement of 3D organoid volume using the PLayer algorithm. (a) Key steps—(I) initialization of n, (II) SME-based weight estimation, (III) layer estimation via overlapping pixels, (IV) Restorer process for image refinement, and (V) layer stacking into 3D volume. (b) Visualization of 3D volume interpolation using the PLayer algorithm, with colors indicating distinct interpolation phases. Reconstructed 3D volumes corresponding to each phase are displayed on the right side of panel (b). Stained with anti-beta III tubulin primary, and a secondary antibody conjugated to 488 nm.
    2D reconstruction of internal organoid layers. Rows (a)–(c) represent layers from n−3 to n+3. Columns 1–4 display the input images, whereas column 5 shows the ground-truth layer (target Layern). Columns 6 and 7 exhibit the SME-based interpolated and reconstructed layers, accompanied by corresponding PSNR and SSIM values. Stained with anti-type IV neurofilament heavy primary, and a secondary antibody conjugated to 488 nm. Panels (d) and (e) provide a comparative analysis of PSNR and SSIM between interpolation methods. Panels (f) and (g) illustrate the results of Restorer-based image reconstruction with different interpolation methods, including GND, cubic, bilinear, and IPF. The x-axis represents the layer number along the axial direction of a randomly selected eight-layer organoid, whereas the y-axis shows SSIM and PSNR values after applying the denoising-based preprocessing step. Panels (h) and (i) offer a statistical comparison of interpolated images and those reconstructed by the Restorer.
    3D reconstruction results of neural organoids. (a) The reference 3D volume. (b) LR 3D volume. (c) Interpolated 3D volume. (d) SR-3D volume without SME. (e) SR-3D volume with SME. Panels (a)–(e) show Z-stack images in 2D space: original high-resolution images (row 1), images without SME (row 2), and images with SME (row 3) across columns labeled 0 to 20 (with one-image gaps of 1 μm)—a red arrow in column 6 highlights how fiber information is preserved in the reconstructed images (rows 2 and 3) compared with the real image (row 1). Panels (f) and (g) depict the 3D volumes of SR-3D without and with SME, respectively. Panel (h) presents the statistical outcomes of 3D reconstruction, evaluated using NIQE. Stained with anti-type IV neurofilament heavy primary and a secondary antibody conjugated to 488 nm.
    Comparative analysis of axial resolution enhancement by factors of 2×, 4×, and 8× utilizing PLayer. The visual representation in the image matrix illustrates the comparative analysis of axial resolution improvements achieved through 2×, 4×, and 8× magnification factors, with the ground truth serving as the reference in the first column. Panels (a) and (b) provide a statistical breakdown of the enhancements observed at each respective compression-magnification factor. Panels (c) and (d) show the reconstruction time required to achieve 2×, 4×, and 8× improvements in axial resolution using either a GPU or a CPU. Stained with anti-type IV neurofilament heavy primary and a secondary antibody conjugated to 488 nm.
    Analyzing the effectiveness of an embedded neural network model in statistical metrics and visualization. (a)–(c) Performance metrics comparison. CNN computed with GPU, CNN computed with Pi, ENN (pruning ratio = 0.2) computed with Pi, and R-ENN computed with Pi. The evaluation metrics used SSIM, PSNR, and reconstruction time. (d)–(f) ENN pruning ratio analysis. Pruning ratios range from 0.4 to 0.8. (g)–(j) Visual results of 3D volume reconstruction. White arrows show the difference between results from R-ENN-Pi, NN-Pi/GPU, and ENN-Pi. Stained with anti-type IV neurofilament heavy primary and a secondary antibody conjugated to 488 nm.
    Tools

    Get Citation

    Copy Citation Text

    Yuanzheng Ma, Davit Khutsishvili, Zihan Zang, Wei Yue, Zhen Guo, Tao Feng, Zitian Wang, Liwei Lin, Shaohua Ma, Xun Guan, "PLayer: a plug-and-play embedded neural system to boost neural organoid 3D reconstruction," Adv. Photon. Nexus 4, 036007 (2025)

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research Articles

    Received: Oct. 17, 2024

    Accepted: Mar. 31, 2025

    Published Online: Apr. 24, 2025

    The Author Email: Ma Shaohua (ma.shaohua@sz.tsinghua.edu.cn), Guan Xun (xun.guan@sz.tsinghua.edu.cn)

    DOI:10.1117/1.APN.4.3.036007

    CSTR:32397.14.1.APN.4.3.036007

    Topics