Abstract
Purpose :
Retinal cysts or fluid spaces accumulated between different layers of eye are key signatures of diseases like age-related macular degeneration (AMD) and macular edema (ME). Accordingly, for accurate screening and disease management, clinicians seek to quantify these lesions using ubiquitous optical coherence tomography (OCT) scans. However, manual delineation is tedious and may induce errors. In response, attempts were made towards automated quantification based on imag processing and machine learning approaches. However, the detection accuracy is still sub-optimal and has scope for improvement. Against this backdrop, we attempt to build a robust fluid segmentation tool leveraging a pre-trained encoder-decoder architecture with residual network (ResNet) at the encoder that was proven to be effective in other segmentation tasks by preserving features across layers.
Methods :
This is a retrospective study consisting of 3175 OCT scans (1048 with fluid/cysts and 2127 without any lesions) taken from wide-field swept-source optical coherence tomography (OCT) device (Carl Zeiss Plex Elite 9000) with a resolution of 12mm×3mm. The presence of fluid was detected based on our previously validated XGBoost algorithm. We employed ResUNet deep-learning architecture which uses residual network at the encoder (see Fig. 1). Masks required to train the model were obtained by manually labeling fluid regions by a trained expert using the ImageJ tool. Dataset is randomly partitioned into training and testing (unseen) sets with a split-ratio of 90:10. Finally, the OCT scans were resized to (256, 256) and the model was trained using Dice coefficient as loss function (epochs: 80, batch size: 16).
Results :
The proposed method achieves an averaged Dice score of 91.47% against ground truth segmentations and outperforms the existing UNet based approaches as well. Fig. 2 depicts representative test OCT images with both manual as well as algorithmic segmentation and also provides volumetric quantification of cysts in 3D volume for a subject.
Conclusions :
The proposed approach achieves segmentation accuracy close to trained optometrists. Next, we are working towards extending this to identify different types of fluids and also perform voxel-level segmentation that enables volumetric quantification of retinal cysts.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.