Abstract
Purpose :
Detection and tracking of retinal edema by OCT images can play a key role in diagnosis and treatment of the disease. We propose a deep learning based method to automatically segment the retinal edema area (REA), pigment epithelial detachment (PED) and subretinal fluid (SRF) from OCT volumes.
Methods :
The AI challenger dataset for segmentation of retinal edema lesions was used[1].The proposed method consists of two steps. (1) Data preprocessing and augmentation. All OCT images are denoised using bilateral filtering. Then a) flipping in the lateral direction of the OCT images was used to simulate the symmetry of right and left eye; b) rotation was used to simulate different inclination of the retina in the OCT images. These processings are applied randomly, and the training data is augmented with a factor of two. (2) Training segmentation model based on improved V-Net. The augmented training data is used to train the V-Net[2] - an improved version of U-Net[3] with residual blocks and dice loss function. In this paper, the pyramid pooling module[4] (PPM) is added into the bottleneck layer of V-Net, which provides additional contextual information for the segmentation task. The “poly” learning rate policy is used where the learning rate is multiplied by (1-iter/max_iter)power with momentum 0.9, power 0.99 and initial learning rate 0.0001.
Results :
Fig. 1 shows the results of the original V-Net and the improved V-Net. For the original V-Net, the Dice similarity coefficient for PED, SRF and REA volume segmentation are 59.2%, 66.3% and 69.5%, and the average Dice similarity coefficient is 65.0%. While for the improved V-Net, the Dice similarity coefficient for PED, SRF and REA volume segmentation are 66.3%, 72.8% and 75.4%, respectively, and the average Dice similarity coefficient is 71.5%.
Conclusions :
Automated segmentation of retinal edema in patients with co-existence of PED, SRF and REA has been achieved. To improve the segmentation performance, we are working for the further improvement of the convolutional neural network, e.g. adding dilation convolution to the encoder of the network, etc.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.