Abstract
Purpose :
Optical coherence tomography angiography (OCTA) scans offer 3D insights into retinal vasculature; however, deeper layers often exhibit projection artifacts that hamper accurate interpretation of the images. Classical projection artifact removal methods, despite their effectiveness, are slow and may rely on lower-level algorithms like layer-segmentation. We present an approach that leverages the output of classical methods as ground truth (GT) for training a deep learning (DL) model, encoding the principles of the classical method into the weights of a Convolutional Neural Network (CNN).
Methods :
We employed a 2D U-Net architecture with a depth of 4, using flow and structure B-scans as inputs, with projection-removed flow B-scans generated using a classical image processing algorithm as GT. A dataset, obtained from PLEX® Elite 9000 (ZEISS, Dublin, CA) was split based on patient IDs and scan patterns. It consisted of 135 eyes for training and 15 eyes for testing. For network training, we used the Adam optimizer along with mean squared error (MSE) as the loss function.
Results :
Our model effectively eliminated projection artifacts in unseen OCTA scans in the hold-out test set, MSE of 3.7±0.8, 5.8±2.1, and 3.7±1.2 for Deep, RPE, and full-retina slabs, with high normalized cross-correlation scores of 0.98, 0.96, and 0.99, respectively. The DL model processes scans 2 to 4 times faster than the classical method, depending on the scan sizes. This efficiency could be further enhanced with batch processing and network optimization. Figure 1 compares classical and DL methods with unprocessed data, highlighting red-overlay flow signals on structure B-scans. Figure 2 shows slabs processed by both, emphasizing the DL model's accuracy in mirroring classical results.
Conclusions :
Our DL method, trained using GT data from a classical image processing method, demonstrates that hand-designed, classical methods can be instrumental in generating offline training data for DL models. This approach accelerates processing and enables selective B-scan processing while maintaining the accuracy of classical methods in removing projection artifacts in OCTA data.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.