Abstract
Purpose :
To perform both qualitative and quantitative analysis of OCT-GAN and Compensation algorithms for shadow removal from Single-frame optical coherence tomography (OCT) B-scans.
Methods :
2,328 OCT B-scans (75x averaging) and 2,328 single-frame B-scans were acquired through the center of the optic nerve head (ONH) using Spectralis OCT (Heidelberg Engineering, Germany) for both eyes of 13 subjects. Extracted features from three perceptual loss networks pre-trained on Imagenet datasets were used with manual segmentations of retinal shadows (where 1 represented a shadow pixel and 0 represented a lack thereof) to train a generative adversarial network (referred to as OCT-GAN) to simultaneously remove retinal shadows and speckle noise. OCT-GAN and OCT compensation were then used on 97 single-frame OCT B-scans from one OCT volume and results were compared qualitatively (for shadow removal and speckle noise) and quantitatively using layerwise pixel intensities (LPI), an indicator for brightness of a specific retinal layer, and the intralayer contrast (ILC), a measure of shadow visibility ranging from 0 (shadow-free) to 1 (strong shadow). ILC was computed in the Retinal Pigment Epithelium (RPE) layer and LPI was computed for the Retinal Nerve Fiber Layer (RNFL), the Ganglion Cell Layer (GCL) + Inner Plexiform Layer (IPL), the Inner Nuclear Layer (INL), and the Outer Plexiform Layer (OPL).
Results :
The mean ILC decreased from 0.40 ± 0.087 to 0.10 ± 0.074 (74.3 ± 19.2%) vs 0.21 ± 0.17 (51.0 ± 36.1%) for OCT-GAN and OCT compensation respectively, indicating shadow removal for both techniques. LPI increased by 4.34 ± 3.66%, 4.58 ± 4.36%, 8.29 ± 6.91%, 0.4 ± 3.80% for OCT-GAN but decreased by 56.0 ± 12.9%, 70.0 ± 7.70%, 74.3 ± 7.30%, 69.7 ± 15.6% for compensation for the RNFL, GCL+IPL, INL, and OPL layers,respectively. OCT-GAN images had visibly less speckle than compensated images and lacked artifacts commonly found in compensated images (inverted shadows, hyper-reflective spots, noise over-amplification at high depth). Both OCT-GAN and compensation were unable to completely remove large retinal shadows.
Conclusions :
OCT-GAN had better shadow removal in the RPE layer and did not suffer from layer dimming commonly observed with compensation. Both algorithms could be considered as pre-processing steps to improve the diagnosis and prognosis of glaucoma from OCT images.
This is a 2021 ARVO Annual Meeting abstract.