Algorithm 2 Training and inference procedure of SDNet.
From: Self-supervised model-informed deep learning for low-SNR SS-OCT domain transformation
TRAINING: Phase 1: • Inputs: B-scans of noise-free dataset corrupted by noise model • Preprocess:         o Corrupt noise-free image by noise model as in Fig. 3.         o Fill missed A-scans with noise as in Fig. 4. • Output: Train SDNet to learn map from preprocessed corrupted noise-free image to original noise-free images. • Update network parameters: Leads to weights Wh. TRAINING: Phase 2: • Input: Low resolution Volume • Output: Coarse denoising version of input B-scan with PCA based on Algorithm 1. • Update network parameters: leads to weights Ws. Mixing weights: Weighted average between network’s parameters of two phases. INFERENCE: • Input: Low-resolution and noisy B-scan         o Preprocess: Fill missed A-scans with noise as in Fig. 4. • Output: Denoised and super-resolution of input noisy B-scan |
