Abstract
Purpose :
Hyper-reflective foci(HRFs) are associated with age-related macular degeneration(AMD), and with visual outcomes in diabetic retinopathy. Automating the HRF segmentation in retinal tissues would be beneficial for feature extraction and overall disease management. This work presents a fully automated approach for HRF segmentation in spectral domain optical coherence tomography(SD-OCT) images.
Methods :
This study includes SD-OCT data acquired at the University of Bonn (Heidelberg Spectralis; image resolution of 786x496 pixels,scan width 30°). The dataset contained 3669 B-scans from 56 patients (10 diabetic macular edema,46 AMD).
A modified U-Net architecture with a densenet-169 backbone was implemented for the semantic segmentation. A point of innovation is the use of a patch-based strategy that ensures the model encoder is trained on enough pixels of interest. To elaborate, as a pre-processing step, we define the retinal pigment epithelium(RPE) layer for each B-scan (using a previously trained model for OCT layer segmentation). An automated dataloader takes the B-scan, HRF ground truth annotation and the RPE denotation as inputs and generates random patches(8x64x64) for each image around the HRF and RPE denoted pixels(location chosen randomly). These patches are then fed into the U-Net encoder, trained with a compound loss function derived from Dice and focal loss. This strategy allows the encoder to train on pathologically important pixels around the inner/outer retina.
Results :
Our trained model performed well,considering the high class-imbalance of the HRF and non-HRF pixels. The current reported HRF segmentation performance in the literature is an average precision(AP) of 0.71. Our baseline AP was similar(AP 0.70) following a similar strategy. Using our patch-based strategy with region of interest selection, the AP improved to 0.75. The AP was 0.69 when trained with a ResNet50 backbone. Overall, the combination of Dice-focal loss generated higher segmentation performance.
Conclusions :
This study presents a pipeline for semantic segmentation of HRFs corresponding to hard exudative and hyperpigmentation. Implementation of a patch-based strategy to improve region of interest selection for the model encoder, and combination of Dice and focal loss improve the current state of the art performance for HRF segmentation.
This is a 2021 ARVO Annual Meeting abstract.