Deep Learning Reconstructs Speckle-Reduced OCT Images Directly From Wavelength Domain

Optical Coherence Tomography (OCT) currently relies on complex processing to create detailed images, often requiring significant computational resources and suffering from image noise. Maryam Viqar, Erdem Sahin, and colleagues at Tampere University and the Bulgarian Academy of Sciences present a new approach that streamlines this process using deep learning. Their method reconstructs high-quality OCT images directly from wavelength data, bypassing the need for complex resampling and significantly reducing computational demands. By sequentially employing two convolutional neural networks, the team effectively removes noise and enhances image quality, demonstrating a substantial improvement in both visual clarity and computational efficiency, and paving the way for future advances in OCT imaging technology.

This work addresses limitations inherent in conventional techniques, which often rely on resampling data and can introduce artifacts or demand significant computational resources. Instead, scientists engineered a deep-learning pipeline consisting of two sequentially applied convolutional neural networks, the Spatial Domain CNN and the Fourier Domain CNN. This innovative approach promises faster, more accurate imaging for biomedical applications.

The process begins by processing degraded images obtained through Fourier transformation, feeding these into the Spatial Domain CNN. This network reconstructs deteriorated structures while simultaneously suppressing unwanted noise, preparing the data for further refinement. Subsequently, the output from the Spatial Domain CNN is processed by the Fourier Domain CNN, which enhances image quality through optimization directly in the Fourier domain. This sequential application of networks allows for targeted noise reduction and image enhancement without complex resampling or calibration procedures. Quantitative and visual assessments demonstrate the ability of this deep-learning pipeline to achieve high-quality reconstructions while reducing computational complexity. This innovation circumvents challenges associated with high-speed OCT systems, which traditionally require meticulous calibration and resampling to maintain imaging rates. This work bypasses the need for resampling techniques, commonly employed in Fourier Domain OCT systems, which can introduce noise, artifacts, and increase computational demands. The team sequentially applied two convolutional neural networks, a Spatial Domain CNN and a Fourier Domain CNN, to directly reconstruct images from wavelength domain data. Experiments demonstrate that the Spatial Domain CNN effectively reconstructs deteriorated morphological structures from highly degraded images obtained via Fourier transformation, simultaneously suppressing unwanted noise.

The subsequent application of the Fourier Domain CNN further enhances image quality through optimization in the Fourier domain. This dual-network approach eliminates the need for complex calibration procedures and associated hardware, offering a streamlined reconstruction process. The research addresses a key limitation of current high-speed OCT systems, where accurate wavenumber linearization is challenging, particularly at A-scan rates exceeding 1. 3 Gigasamples per second. By removing the reliance on resampling and calibration, the team’s method avoids potential issues with clocking glitches and inaccuracies. Furthermore, the deep learning approach mitigates speckle noise, a common artifact in OCT imaging, without requiring extensive averaging that can blur image details or be ineffective with dynamic samples. The team successfully demonstrated a method that directly reconstructs speckle-reduced OCT images from wavelength domain data, utilizing sequential Spatial Domain and Fourier Domain Convolutional Neural Networks. By processing data in this manner, the system avoids the need for resampling into the wavenumber domain, simplifying the reconstruction process and reducing computational demands. The developed method achieves high-quality image reconstruction while simultaneously suppressing speckle noise, a common issue in low-coherence interferometry.

Comparative analysis confirms the efficacy of this deep-learning approach, demonstrating performance comparable to existing commercial OCT systems. This work establishes a framework for future innovation in OCT image reconstruction, offering a streamlined and efficient alternative to conventional methods. Researchers acknowledge that the performance of their method is dependent on the quality of the training data and the specific characteristics of the light source used in the OCT system. Future research directions include exploring the application of this approach to different OCT modalities and investigating the potential for further optimization of the neural network architecture.