An automated choroid segmentation approach using transfer learning and encoder-decoder networks

Abstract

Purpose : Deep neural networks have been used for choroidal segmentation using OCT images. However, the complexities of choroidal structure require the use of a large amount of data and computational time for training a model. We hypothesize that integration of transfer learning to an encoder-decoder network can enable latent features to be detected under the constraint of limited data. We tested this hypothesis with a computational framework that adapts geometrical transformations, spatially adaptive augmentation, and variance scaling weight initialization.

Methods : We used 120 OCT image-mask pairs of the choroid and randomly assigned 96 of them for training and 24 for validation. We augmented the training set to 1248 images using rotation and flipping. All images were resized to 128×128 pixels using bilinear interpolation and then masks were binarized. We built U-Net and UNet++ (encoder-decoder network) models on a VGG-19 backbone for allowing pretrained (PT) ImageNet weights in the encoder. We adapted a Glorot uniform initialization otherwise. We experimented with freezing (FR) the first 30% of the backbone layers when using transfer learning while using the final layer outputs for U-Net and UNet++. The models were trained for 10 epochs using Dice-score loss as a quantitative measure, and a model performance was evaluated using an average precision as a qualitative measure.

Results : We obtained average precision scores of 0.92 for U-Net and UNet++. The integration of transfer learning yielded an average increase of 2% in U-Net and 4% in UNet++. A subjective (visual) evaluation confirmed that the segmented choroidal regions are meaningful (Fig. 1) with respect to the increase in precision values. Freezing of the initial layers slightly reduced the precision value.

Conclusions : The transfer learning approach can significantly help to improve encoder-decoder networks for choroidal segmentation in OCT scans. Our computational framework comprises U-Net and UNet++ on a VGG-19 backbone and uses the PT encoder to perform choroidal segmentation. Our future research will include more rigorous subjective evaluation with choroidal OCT images of unhealthy eyes to develop a fully automated choroidal segmentation system.

This is a 2021 ARVO Annual Meeting abstract.

 

Fig. 1: Top row: Original, Segmentation, Mask; Middle row: U-Net Results - without PT, with PT, and with PT and FR; Bottom row: UNet++ Results without PT, PT, and with PT and FR.