Accurate attenuation characterization in optical coherence tomography using multi-reference phantoms and deep learning.

The optical attenuation coefficient (AC), a crucial tissue parameter indicating the rate of light attenuation within a medium, enables quantitative analysis of tissue properties and facilitates tissue differentiation. Despite its growing clinical significance, accurate quantification of AC from optical coherence tomography (OCT) signals remains a pressing concern. This study comprehensively investigates the factors influencing the accuracy of quantitative AC extraction among existing OCT-based AC extraction algorithms. Subsequently, we propose an approach, the Multi-Reference Phantom Driven Network (MR-Net), which leverages multi-reference phantoms and deep learning to implicitly model factors affecting OCT signal propagation, thereby automatically regressing AC. Using a dataset from Intralipid and silicone-TiO₂ phantoms with known AC values obtained from a collimated transmission system and imaged with a 1300 nm swept-source OCT system, we conducted a thorough comparison focusing on data length, out-of-focus distance, and reference phantoms' attenuation among existing OCT-based AC extraction algorithms. By leveraging this extensive dataset, MR-Net can automatically model the complex physical effects in the transmission process of OCT signals, significantly enhancing the accuracy of AC predictions. MR-Net outperforms other algorithms in all metrics, achieving an average relative error of only 10.43% for calculating attenuation samples, significantly lower than the lowest value of 23.72% achieved by other algorithms. This method offers a quantitative framework for disease diagnosis, ultimately contributing to more accurate and effective tissue characterization in clinical settings.