Neural Surrogate Modeling for Optical Coherence Tomography Reconstruction with Real-Time Constraints

Abstract

Optical Coherence Tomography (OCT) has established itself as a paramount imaging modality in ophthalmology, cardiology, and dermatology, offering non-invasive, high-resolution cross-sectional visualization of biological tissues. However, the computational burden associated with high-fidelity image reconstruction—particularly when employing iterative Compressed Sensing (CS) algorithms or complex dispersion compensation techniques—often precludes real-time application in time-sensitive clinical environments such as intraoperative surgical guidance. This paper introduces a novel Neural Surrogate Modeling framework designed to approximate the complex inverse scattering physics of OCT reconstruction while strictly adhering to real-time latency constraints. By leveraging a hardware-aware deep learning architecture, specifically a lightweight Fourier-domain convolutional neural network optimized via neural architecture search, we successfully map raw interferometric data directly to structural images, bypassing the latency of traditional iterative solvers. We introduce a multi-objective loss function that balances structural fidelity, perceptual quality, and sparsity constraints. Furthermore, we provide a comprehensive analysis of model quantization and tensor acceleration techniques necessary to deploy these models on edge-computing devices. Our results demonstrate that the proposed neural surrogate achieves reconstruction quality competitive with state-of-the-art iterative methods (PSNR > 32 dB) while operating at inference speeds exceeding 150 frames per second on standard GPU hardware, effectively bridging the gap between high-fidelity imaging and real-time feedback loops.