Physics-Guided and Data-Driven Conditional GAN Data Augmentation for Depth-Resolved Spatial Frequency Domain Imaging Optical Property Estimation: A Simulation-Based Comparative Study

Background: Spatial frequency domain imaging (SFDI) is a non-contact, wide-field optical imaging technique that quantifies tissue absorption (μₐ) and reduced scattering coefficients (μs′). Conventional SFDI provides depth-averaged 2D maps. Methods: This study presents a simulation-based framework for depth-resolved optical property estimation over five depth layers, including discrete 3D inclusions. The framework also includes a comparison of two conditional GAN (cGAN) designs for synthetic data augmentation. Monte Carlo simulation generated 1000 heterogeneous tissue samples, with μₐ and μs′ varied independently. Two cGANs with identical architectures were trained. The first was a physics-guided (PG-cGAN), with six physics-based loss terms; the second was a data-driven (DD-cGAN) with no physics losses. Each cGAN generated 5000 synthetic samples. A multi-output U-Net was trained on each synthetic dataset separately and evaluated on the same held-out set of 200 real MCX samples to assess downstream optical property estimation accuracy. Results: Synthetic augmentation significantly outperformed the real-only baseline, with DD-cGAN reducing μₐ error by 47.7% and μs′ reducing error by 34.6%, while PG-cGAN reduced μₐ error by 43.2% and μs′ error by 33.5% (Exp4: μₐ R2 = 0.9953, μs′ R2 = 0.9328; Exp9: μₐ R2 = 0.9957, μs′ R2 = 0.9325; ΔR2 < 0.002). Per-slice analysis showed that Exp 9 achieved lower μₐ in deep slices, whereas Exp 4 had advantages in superficial slices. The two synthetic datasets also demonstrated complementary strengths. DD-cGAN produced samples that are closer to the physics targets (DC–μs′ error: 4.9% vs. 28.4%), whereas the PG-cGAN achieved higher image fidelity (paired SSIM: 0.622 vs. 0.599). Both designs achieved equivalent cross-condition diversity (0.789 vs. 0.791). Independent validation on a different MCX dataset (500 samples, seed 2026) confirmed that the derived physics relationships were consistent within the simulated parameter range. Conclusions: Both cGANs designs demonstrate comparable downstream performance. DD-cGAN provides improved physical correlation fidelity, whereas PG-cGAN provides improved image fidelity. These comparisons have been demonstrated using in-distribution MCX test data. Clinical translation will require experimental validation on phantoms and ex vivo tissue samples.