Direct prediction of retinal layer boundaries in OCT images

Diabetic macular edema (DME) is the leading cause among those having diabetes and is caused by leaking blood vessels resulting in accumulated fluid pools in the retina. It can be cured if treated in time, however, its diagnosis requires the semantic segmentation of the retina, which is a time-consuming, error-prone procedure usually done by domain experts, giving rise to the desire to automate the task. Current state-of-the-art deep-learning-based approaches do the layer segmentation by segmenting the whole B-scan pixel-wise. These approaches cannot be directly optimized for layer boundary prediction, the predictions for the layer boundary coordinates are computed in a subsequent step. In this thesis I examine a family of approaches that directly predicts the coordinates of the retinal layer boundaries, thus the model can be optimized for coordinate prediction. The proposed approaches are intuitively based on the way how domain experts do manually the segmentation: they do not classify each pixel individually – as related works do – but they rather localize the boundaries in certain areas of the image and continue the curves by sequentially segmenting the neighbor areas. The proposed methods accomplish the same by first segmenting the image columnwise and smoothing these segmentations with an LSTM cell or 1D convolutions. Further, in order to compare the performance of the proposed approaches I reimplement two of the state-of-the-art approaches for retinal layer segmentation, ReLayNet and DeepRetina. The proposed method achieves a lowest contour error of 1.92 on the Duke dataset, outperforming the reproduced version of ReLayNet (2.20) and achieving comparable performance to the reproduced version of DeepRetina (1.75) with 90% fewer learnable parameters. The source code is publicly available on https://github.com/loerinczy/master-thesis.