Abstract
Purpose :
Modern SS-OCT systems with MHz A-scan rates allow real-time volumetric visualization, but suffer from increased noise levels due to the high acquisition speed. In this work, we applied a light-weight neural network for noise reduction in a microscope-integrated 4D-OCT system while preserving its real-time capability.
Methods :
Imaging is performed with a 1060 nm SS-OCT prototype with an A-scan rate of 1.2 MHz. Its sample arm is integrated into a surgical microscope (ARTEVO 800®, ZEISS, Jena, Germany) using an add-on module. The system is capable of 4D imaging using a spiral scanning trajectory.
The denoising network, a light-weight U-Net variant, was trained on unpaired B-scans from the prototype as well as a PLEX® Elite 9000 (ZEISS, Dublin, CA), using self-supervised learning based on structured Noise2Void.
In order to allow real-time denoising, the network is incorporated into the reconstruction pipeline, executed on a GPU (Nvidia Titan RTX), using Nvidia TensorRT. Processing is limited to two-dimensional buffers with an effective size of 704×4704 pixels. A single volume consists of 25 such buffers, and the volume rate is 10 volumes per second. For denoising, these buffers are split into pseudo-batches with a size of 4 and then processed. Once a whole volume is acquired, rendering is performed on a second, identical GPU using the software CAMPVis.
Results :
With our approach and system, volumes with a diameter of 4.5 mm can be denoised and displayed at 10 volumes per second. An example of volume renderings of a porcine retina, showing the effect of denoising on the displayed volumes, is shown in Figure 1.
Major differences are highlighted by arrows: Denoising reduces obscuring noise (orange) in the vitreous. Furthermore, tissue structures and surfaces (blue) appear smoother due to reduction of speckle, which leads to coarse surfaces otherwise. Meanwhile, due to the two-dimensional processing approach, additional latency is kept below the available per-buffer processing time of 4 ms.
Conclusions :
Live denoising using a real-time capable, light-weight network can improve volumetric visualization in 4D-OCT by reducing the relatively large amount of noise. With the presented approach, only a few milliseconds of latency for each volume are added, preserving the potential intra-surgical applicability while improving the imaging quality.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.