Abstract
Purpose :
To develop artificial intelligence (deep learning) algorithms to boost (denoise and deshadow) optical coherence tomography images with applications to the optic nerve head (ONH).
Methods :
Volume scans consisting of 97 horizontal B-scans were acquired through the center of the ONH using a commercial OCT device (Spectralis) for both eyes of 20 subjects. For each eye, single-frame (without signal averaging), and multi-frame (75x signal averaging) volume scans were obtained. A total of 3,880 B-scans (either single- or multi-frame) were split into training (70%) and testing (30%) sets. To boost OCT image quality, we used 2 approaches. For the first, we used a custom deep learning network trained with clean B-scans (multi-frame B-scans), and their corresponding noisy B-scans (clean B-scans + gaussian noise) to denoise single-frame B-scans. The performance of the denoising algorithm was assessed qualitatively, and quantitatively on unseen images using the signal-to-noise ratio (SNR). For the second approach, we used a generative adversarial network (GAN; incorporating modified U-Net architectures) to automatically detect and remove shadows according to a predicted ‘shadow score’. This latter was produced from a network trained on B-scans with manually segmented shadows (in imageJ-Fiji). The performance of the deshadowing algorithm was assessed qualitatively, and quantitatively on unseen images using the intra-layer-contrast (a measure of shadow removal that varies between 0 [shadow-free] and 1 [strong shadow]).
Results :
The denoising algorithm successfully denoised unseen single-frame OCT B-scans. The denoised B-scans were qualitatively similar to their corresponding multi-frame B-scans, with enhanced visibility of the ONH tissues (Figure 1). The mean SNR increased from 4.02 ± 0.68 dB (single-frame) to 8.14 ± 1.03 dB (denoised). The deshadowing algorithm successfully removed shadows from unseen multi-frame OCT B-scans (Figure 1). The mean intralayer contrast decreased from 0.28 ± 0.14 (shadowed B-scan) to 0.04 ± 0.05 (deshadowed).
Conclusions :
We have proposed novel deep learning algorithms to boost the quality of existing OCT images. Our work offers the possibility of producing low-resolution, low-quality OCT hardware complemented with artificial intelligence software technology to achieve high-image quality for a fraction of existing OCT device market cost.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.