Logo image
Back
Synthetic-data-driven optical-property mapping in OCT-like phantoms using an inclusion-guided consistency-regularized U-Net
Journal article   Open access   Peer reviewed

Synthetic-data-driven optical-property mapping in OCT-like phantoms using an inclusion-guided consistency-regularized U-Net

Tahasin Ahmed Fahim and Fatema Binte Alam
Optics continuum, Vol.5(9), pp.2914-2929
2026
DOI: https://doi.org/10.1364/OPTCON.607241
pdf
Published5.92 MBDownloadView
Open Access CC BY V4.0

Abstract

Optical coherence tomography (OCT) provides depth-resolved structural information, but estimating dense optical-property maps from intensity data is an ill-conditioned inverse problem. This study presents LA-PG-UNet (localization-aware, prior-guided U-Net), an inclusion-guided and output-consistency-regularized framework evaluated entirely on controlled synthetic OCT-like phantoms. The network estimates the absorption coefficient μa, the reduced scattering coefficient μs′, and a study-defined composite attenuation coefficient μc = μa + μs′ from a single-channel OCT-like B-scan produced by a phenomenological attenuation, surface-response, roll-off, speckle, and noise generator. A multi-phantom dataset of 1500 samples spans smooth layered, rough multilayer, heterogeneous inclusion, and shifted-parameter noisy conditions. Against a GroupNorm U-Net baseline, LA-PG-UNet reduced inclusion-region μc mean absolute error (MAE) in all four test partitions (13.63%, 6.80%, 4.43%, 5.36%; all significant after Holm correction) and reduced output-consistency error in all four (significant in three), while the baseline retained lower global μc MAE—a genuine trade-off between background-dominated full-image reconstruction and localized inclusion recovery. Regional errors for all three coefficients show the inclusion gain is carried almost entirely by μs′, with μa unchanged. A controlled component-wise ablation under shared initialization shows that the output-consistency loss, the inclusion-weighted regression term, the auxiliary branch, and the feature gate each contribute measurably and in different directions, and pixel-level AUROC of 0.74–0.84 indicates coarse but non-random inclusion localization despite low dice. A training-calibrated classical depth-resolved attenuation estimator and a two-output analytical comparator were both outperformed in inclusion-region μc MAE. All results describe generalization within the synthetic simulator; no experimental OCT or physical-phantom validation is claimed.

Details

Metrics

1 Record Views
Logo image