Abstract
Purpose :
Adaptive optics has enabled in vivo imaging of the living human retina with diffraction limited spatial resolution and thereby can image the retinal structure at the cellular level. However, adaptive optics may not readily be available for all imaging modalities, and in some cases its performance may be compromised by incomplete compensation of ocular wave aberrations. We investigate a deep learning based method to enhance the spatial resolution of retinal images without or with adaptive optics.
Methods :
Twelve high resolution retinal images were obtained using an adaptive optics near-confocal ophthalmoscope (AONCO) as the ground truth. To model the image blur induced by ocular optical defects, we introduced various wave aberrations (with varying Zernike coefficients up to the 7th order). To expand the dataset, the AONCO images were split into small patches (2820 patches/image on average). With 8 different blurring kernels applied, there were 270,736 image patches in total for training and testing. In application of the trained networks, the corrected patches were combined to form images of their original sizes.
The artificial neural network was based on a U-net architecture. We performed a 4-fold cross validation study with a quarter of the images reserved for testing per cross validation.
The normalized mean squared error (NMSE) and structural similarity index measure (SSIM) were calculated for the images before and after correcting for the blurring effects, using the ground truth images as the reference.
Results :
After correction with the neural networks, the NMSE was reduced by 36% on average, from 0.063 ± 0.026 to 0.040 ± 0.013. The SSIM was improved by 89% on average, from 0.22 ± 0.10 to 0.40 ± 0.14. The improvements were significant (by paired t-tests, p< 0.001 for both the NMSE and the SSIM). Image examples are shown in Figures 1 and 2.
Conclusions :
Our preliminary results showed that deep learning may be useful in correcting the retinal image blur caused by ocular wave aberrations. Further optimization with larger datasets and more types of blurring effects are underway to improve its performance and the generalizability of the network.
This is a 2021 ARVO Annual Meeting abstract.