Enhancing depth profile reconstruction in FD-OCT using ResLat-GAN

We propose a residual latent generative adversarial network (ResLat-GAN), a conditional GAN (cGAN) architecture designed to suppress artifact in one-dimensional (1D) fourier domain optical coherence tomography (FD-OCT) signals. The model integrates residual encoding and latent-space-aware modules to enhance structural fidelity in both single-layer and double-layer depth profiles. Trained on simulated FD-OCT data with paired artifact-free references, ResLat-GAN effectively transforms noisy profiles into structurally accurate reconstructions, outperforming baseline models such as Pix2Pix GAN, CycleGAN, and Autoencoder. It reaches structural accuracies of 57.6% and 38.4% for single-layer and double-layer profiles, respectively, demonstrating more than a twofold improvement over baseline methods. These results underscore the model’s effectiveness in reducing artifacts and preserving critical structural features in complex FD-OCT profiles.