Advanced Imaging, Volume. 2, Issue 5, 055001(2025)

3D Gaussian adaptive reconstruction for Fourier light-field microscopy On the Cover

Chenyu Xu, Zhouyu Jin, Chengkang Shen, Hao Zhu, Zhan Ma, Bo Xiong*, You Zhou*, Xun Cao, and Ning Gu
Figures & Tables(4)
Principle of 3D Gaussian adaptive tomography (3DGAT). (a) The optical setup and forward imaging process of the FLFM system. (b) The schematic of the training pipeline of 3DGAT. NOP, native object plane; OBJ, objective lens; TL, tube lens; NIP, native image plane; FL, Fourier lens; MLA, microlens array; CAM, camera.
Performance and resolution evaluation of 3DGAT on synthetic data. (a) Maximum intensity projections (MIPs) of synthetic fluorescent beads restored by RL deconvolution and 3DGAT, compared to the ground truth. (b), (c) Intensity profile comparisons of RL deconvolution (gray) and 3DGAT (red) with the ground truth (black) along the yellow dashed lines in (a). (d) MIPs of synthetic reticular structures reconstructed by RL deconvolution and 3DGAT, alongside the ground truth. (e) Intensity profile comparisons of RL deconvolution (gray) and 3DGAT (red) with the ground truth (black) along the yellow dashed line in (d). (f) Quantitative evaluation using PSNR, SSIM, and LPIPS metrics for RL deconvolution and 3DGAT across the depth range. (g), (h) x–y and x–z MIP images of synthetic fluorescent lines with varying intervals in the LF central view, as well as reconstructions by RL deconvolution, 3DGAT, and the ground truth. (i), (j) Intensity profile comparisons of the LF central view (gray), RL deconvolution (pink), and 3DGAT (red) with the ground truth (black) along the yellow dashed line in (g), (h). A.U., arbitrary units; Scale bar, (a), (d) 10 µm.
Comparison between RL deconvolution and 3DGAT with different loss functions. (a) MIPs of the simulated dandelion sample restored by RL deconvolution and 3DGAT with different losses, such as MSE, MAE + SSIM, and MSE + FDL, compared with the ground truth. (b), (c) Enlarged views of the regions outlined by the dashed boxes in corresponding colors in (a). (d) Normalized intensity profiles of the ground truth (gray), RL deconvolution (brown), and 3DGAT with MSE loss (purple), MAE + SSIM loss (pink), and MSE + FDL loss (red) along the white dashed line in (b). (e) Quantitative metrics for reconstruction evaluation across the depth range. Scale bar: (a) 100 µm; (b), (c) 20 µm.
Reconstruction of experimentally captured zebrafish data. (a) Raw Fourier light-field image of zebrafish data acquired from Ref. [28]. (b) x–y MIP images of results obtained by RL deconvolution, raw 3DGAT, and effective-rank-regularized 3DGAT (3DGAT-erank). (c) Enlarged views of the white and green dashed boxes in (b). (d) Fourier domain visualization of x–y and x–z MIP images recovered by the corresponding methods. (e) Normalized intensity profiles along the green dashed line in (c). (f) Fourier ring correlation quality estimate (FRC-QE) scores of three methods, respectively. Scale bar: (a), (b) 50 µm; (c) 30 µm.
Tools

Get Citation

Copy Citation Text

Chenyu Xu, Zhouyu Jin, Chengkang Shen, Hao Zhu, Zhan Ma, Bo Xiong, You Zhou, Xun Cao, Ning Gu, "3D Gaussian adaptive reconstruction for Fourier light-field microscopy," Adv. Imaging 2, 055001 (2025)

Download Citation

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

Category: Letter

Received: Apr. 1, 2025

Accepted: Aug. 21, 2025

Published Online: Feb. 28, 2025

The Author Email: Bo Xiong (xiongbo@pku.edu.cn), You Zhou (zhouyou@nju.edu.cn)

DOI:10.3788/AI.2025.50001

Topics