Abstract
Purpose :
Retinal layer segmentation deep learning (DL) models trained on data from a source OCT device usually do not generalize to another target device due to the differences in visual appearance. The alternative of labeling OCT images from the target device is time consuming and costly. This study evaluates a semi-supervised contrastive learning method that requires no labeled images from the target device, which can support more efficient development of DL models across multiple OCT devices.
Methods :
SegCLR is a semi-supervised contrastive learning method with proven success in retinal fluid segmentation where supervised learning and contrastive learning are performed simultaneously. SegCLR was evaluated in a layer segmentation problem (source device: Heidelberg Spectralis; target device: ZEISS Cirrus). The data sets were obtained from 2 clinical trials evaluating neovascular age-related macular degeneration: Avenue (NCT02484690; Spectralis) and Harbor (NCT00891735; Cirrus). Inner limiting membrane, outer plexiform layer, retinal pigment epithelium, and Bruch’s membrane were manually annotated by expert graders in a subset of both trials (Table 1). The performance of SegCLR was compared against a baseline (BL) model trained on labeled data from Avenue, and an upper bound (UB) model trained on labeled data from Harbor. Labeled Harbor data were only used for evaluations and training of the UB model (Fig 1). The data split was done at the patient level. The Dice coefficient on Harbor test set averaged over the 4 layers is reported as a performance metric.
Results :
The mean Dice coefficient (standard deviation [SD] across 5 replicates) in the Harbor test set was 70.15 (15.97) for the BL model, 83.77 (0.19) for the UB model, and 83.16 (0.75) for the SegCLR model. SegCLR improved the performance by more than 13% from the BL model and performed similarly to the UB model in mean Dice coefficient. Comparing SD, SegCLR largely gained stability from the BL model.
Conclusions :
SegCLR shows the ability to learn from unlabeled target data in OCT layer segmentation, enabling the model to quickly and accurately adapt to new OCT devices for retinal layer segmentation in the future, which can contribute to bringing efficiency and reliability to the clinical assessment of OCT images.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.