Denoising of surgical microscope-integrated optical coherence tomography images

Abstract

Purpose : Application of multi-frame averaging for OCT noise reduction is limited in intra-surgical application. Meanwhile, the data rate in modern systems, capable of acquiring B-scans covering the entire eye length, is challenging for other noise reduction algorithms. We present real-time capable visualization of the entire eye-length with neural-network-based denoising.

Methods : Data were acquired using a 1060 nm SS-OCT prototype at an A-scan rate of 100 kHz. The network is a U-Net with an adjustable number of filters per convolution layer at the highest level, which is doubled after each of the two down-sampling steps. It is trained with adaptations of structured Noise2Void for live visualization and Noise2Noise for offline denoising. The training set consists of B-scans from the prototype and a PLEX^® Elite 9000 (ZEISS, Dublin, CA). For real-time denoising, the network is incorporated directly into the reconstruction pipeline of the system. Live B-scans have a size of 6656×1300 pixels and are processed at full size. The native frame rate of the prototype is 40 Hz. Only for display, these B-scans are rescaled to 1000×800 pixels. The full volumetric scan of a porcine eye consists of 499 B-scans with 13312×500 pixels each.

Results : Example screenshots from live B-scans are shown in Figure 1, including achieved frame rates. All tested network complexities up to n=16 would allow video rate display (≥24 Hz). However, the largest of these complexities almost halves the frame rate from the native 40 Hz, while it is preserved for n∈{4,8}. The effect of denoising on a volumetric scan covering the entire eye length is shown in Figure 2. Denoising qualitatively removes noise and speckle in both scenarios. Furthermore, weak signals such as the artificial lens, the posterior capsular bag and several retinal layers, and fine details, for example on the corneal surface, are better visible both for the anterior and the posterior segment of the eye.

Conclusions : We have shown that full eye-length B-scans with a size comparable to or beyond 4K resolution can be denoised in real time. Live image quality, but even more so offline rendered volumetric images, benefit from denoising by sharper details and higher qualitative contrast.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1: Live B-scan display of a model eye with and without denoising.

 

Figure 2: Full eye-length volume rendering of a porcine eye after simulated cataract surgery.