Fig. 3.

Real-time self-fusion neural network processing pipeline. Three OCT B-scans were acquired and loaded into LibTorch-GPU tensors. The LibTorch-based OCT processing that included background subtraction (-BG), Hanning spectral window (Hann), dispersion-compensation (Disp), Fourier transform (FFT), logarithmic compression (Log), and image cropping (Crop). The input images for self-fusion neural network were then normalized (Norm), motion-corrected (MC), processed with the pretrained self-fusion neural network (SFNN), and contrast-adjusted (CA) before being displayed (ID).