Optical coherence tomography (OCT) now relies on complex processes to create detailed images, often suffering from important computational resources and image noise. Maryam Viqar, Erdem Sahin, and colleagues at Tampere University and the Bulgarian Academy of Sciences present new approaches to streamline this process using deep learning. These methods directly reconstruct high-quality OCT images from wavelength data, bypassing the need for complex resampling and significantly reduce computational demand. By adopting two convolutional neural networks in sequence, the team effectively removes noise, improves image quality, shows significant improvements in both visual clarity and computational efficiency, paving the way for future advancements in OCT imaging technology.
This work addresses limitations inherent to traditional techniques. This often relies on data resampling, allowing you to introduce artifacts and request important computational resources. Instead, scientists designed a deep learning pipeline consisting of two consecutively applied convolutional neural networks, a spatial domain CNN and a Fourier domain CNN. This innovative approach promises faster and more accurate imaging for biomedical applications.
The process begins by processing the degraded images obtained by the Fourier transform and feeding them to the spatial domain CNN. This network reconstructs degraded structures, simultaneously suppresses unnecessary noise and prepares data for further improvement. The output from the spatial domain CNN is then processed by the Fourier domain CNN to improve image quality through Fourier domain optimization. This sequential application in the network allows for targeted noise reduction and image enhancement without complex resampling or calibration steps. Quantitative and visual assessments demonstrate the ability of this deep learning pipeline to reduce computational complexity while achieving high-quality reconstruction. This innovation avoids the challenges associated with high-speed OCT systems. This traditionally requires meticulous calibration and resampling to maintain the imaging rate. This work bypasses the need for resampling techniques commonly employed in Fourier domain OCT systems, which can introduce noise, artifacts and increase computational demand. The team sequentially applied two convolutional neural networks, spatial domain CNN and Fourier domain CNN, to directly reconstruct images from wavelength domain data. Experiments show that spatial domain CNNs effectively reconstruct degraded morphological structures from highly degraded images obtained via Fourier transform, simultaneously suppressing unnecessary noise.
Subsequent applications of Fourier Domain CNN further improve image quality through Fourier Domain optimization. This dual-network approach eliminates complex calibration procedures and the associated hardware needs and provides a streamlined reconfiguration process. This study addresses the important limitations of current high-speed OCT systems where accurate wavenumber linearization is difficult, especially at A-scan rates above 1.3 gigasamples per second. The team's method avoids potential problems recording glitches and inaccuracies by removing responsibilities on resampling and calibration. Furthermore, the deep learning approach reduces speckle noise, a common artifact in OCT imaging. This does not blur image details or have no effect on dynamic samples. The team successfully demonstrated how to utilize convolutional neural networks of sequential spatial and Fourier domains to directly reconstruct speckle-reduced OCT images from wavelength domain data. By processing the data in this way, the system avoids the need to resample into the WaveNumber domain, simplifies the reconfiguration process and reduces computational demand. The developed method achieves high quality image reconstruction while simultaneously suppressing speckle noise, a common problem with low coherence interferometry.
Comparative analysis confirms the effectiveness of this deep learning approach and demonstrates performance comparable to existing commercial OCT systems. This work establishes a framework for future innovation in the reconstruction of OCT images and provides a reasonable and efficient alternative to traditional methods. Researchers acknowledge that the performance of the method depends on the quality of the training data and the specific characteristics of the light source used in the OCT system. Future research orientations include investigating the application of this approach to various OCT modalities and examining the possibilities for further optimization of neural network architectures.
👉Details
🗞 Reconstructing optical coherence tomographic images from wavelength space using deep learning.
🧠arxiv: https://arxiv.org/abs/2509.18783
