Abstract
Purpose :
Optical coherence tomography (OCT) imaging has benefitted ophthalmic diagnostics by enabling depth-resolved volumetric imaging of ocular structures. Inherent speckle noise degrades the OCT image quality fficulting the identification of anatomical features and pathologic features. Frame-averaging is a ubiquitous method for increasing signal-to-noise ratio (SNR), however, the need to acquire multiple repeated frames increases the imaging time, introduces motion artifacts, and adds to potential patient discomfort. We recently introduced a method called self-fusion, which reduces speckle noise and enhances OCT SNR using adjacent frames, thus reducing the need to repeated image the same location and extending imaging time. Uniquely, self-fusion integrates image similarity metrics between adjacent frames to maintain lateral resolution. We have trained a convolutional neural network to offset the computational overhead of self-fusion and, here, present a framework for performing self-fusion in real-time.
Methods :
The neural network was designed in PyTorch based on the U-net architecture. The training and validation sets were first processed with using self-fusion to yield ground truth images and then both the noisy and self-fused images were fed into the training algorithm to fit the model and optimize the hyperparameters (Fig. 1A). An independent test set was used for unbiased evaluation of a final model. Since the data acquisition software (DAQ) was written in C++, the pretrained model was exported to a file to interface with the DAQ (Figure 1B). A self-fusion network that was pretrained to fuse 3 frames was implemented to further increase display rates.
Results :
Figure 1C shows a noisy OCT B-scan, a 7-frame averaged image, and a 3-frame self-fused image. There is a clear gain in peak SNR in the 3-frame self-fused image (20.2 dB) over both the raw and 7-frame averaged image (13.8 dB).
Conclusions :
This approach delivers a fast and robust OCT denoising alternative to frame-averaging without the need for repeated image acquisition. Real-time self-fusion image enhancement will enable improved localization of OCT field-of-view relative to features-of-interest and improved sensitivity for anatomic features of disease.
This is a 2021 ARVO Annual Meeting abstract.