Fast and Compressed ADA-UNet: Refractive-Index Reconstruction with Latent-Domain Low-Rank Compression

Authors: Chiang, Po-Jui; Lin, Yu-Da; Yang, Cheng-Hong

Source journal: IEEE transactions on image processing : a publication of the IEEE Signal Processing Society (2026-11-17)

DOI: 10.1109/TIP.2026.3752834

Abstract

Optical diffraction tomography (ODT) provides label-free three-dimensional refractive-index (RI) maps, but conventional reconstructions are highly sensitive to speckle noise and severely under-sampled illumination angles. We propose ADA-UNet, a compact two-stage framework in which (i) a modality-specific autoencoder learns a compact representation of stitched diffraction inputs, and (ii) a latent-domain ResUNet maps these interferometric latent codes to RI latent codes for non-iterative reconstruction. For compactness, the interferometry-to-RI mapping is expressed in Dirac form, but the network itself is implemented entirely with standard convolutional operators on GPU hardware. To reduce computational cost without changing the reconstruction objective, low-rank factorization is applied specifically to the ADA-aligned latent mapper. The resulting compressed model, ADA-UNetc, reduces the parameter count from 12.5M to 6.0M (a 52% reduction) while maintaining nearly unchanged reconstruction quality, with SSIM changing from 0.943 to 0.940 and NRMSE from 0.051 to 0.052. Across all evaluated ODT, cross-modal image-synthesis, and adaptation baselines, the proposed model achieves the best overall fidelity and a measured per-volume inference time of 187 s, compared with 221-840 s for the baselines. Experiments on synthetic and real 2-D/3-D data further demonstrate robustness under evaluation-time noise perturbations and sparse-angle acquisition, highlighting the potential of ADA-UNet for efficient and reliable quantitative phase imaging.

Read the full text

Related topics

Related open access articles

Abstract and bibliographic data from PubMed (U.S. National Library of Medicine). Copyright of the abstract belongs to the original publisher or authors.