A Deep Learning Framework for Segmentation of Retinal Layers from OCT Images

Abstract

Segmentation of retinal layers from Optical Coherence Tomography (OCT) volumes is a fundamental problem for any computer aided diagnostic algorithm development. This requires preprocessing steps such as denoising, region of interest extraction, flattening and edge detection all of which involve separate parameter tuning. In this paper, we explore deep learning techniques to automate all these steps and handle the presence/absence of pathologies. A model is proposed consisting of a combination of Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM). The CNN is used to extract layers of interest image and extract the edges, while the LSTM is used to trace the layer boundary. This model is trained on a mixture of normal and AMD cases using minimal data. Validation results on three public datasets show that the pixel-wise mean absolute error obtained with our system is 1.30±0.48 which is lower than the inter-marker error of 1.79 ± 0.76. Our model's performance is also on par with the existing methods.

Automatic glaucoma classification using color fundus images based on convolutional neural networks and transfer learning

“…The assessment of deep learning algorithms with transfer learning has also been addressed in [43][44][45] implementing studies with greater number of images than previous works and achieving expert level accuracy and high sensitivity and specificity. Finally, in OCT there are also recent studies applying CNNs for glaucoma detection [46] or segmentation of layers [47,48]. In Table 1 we present a summary of the methods used for glaucoma detection, describing the data sets used and the results reported.…”

Automatic Segmentation of Macular Holes in Optical Coherence Tomography Images: A review

“…González-López [105] created an algorithm tolerant to noisy scenarios. Gopinath [108] developed a method that not requires any preprocessing step. Duan [110] successfully segmented the OCT image that contains some broken retinal layers.…”

“…Authors Year Preprocessing Segmentation Classification Nasrulloh et al [11] 2018 Yes Yes No Keller et al [26] 2016 Yes Yes No Miri et al [96] 2016 Yes Yes No Zhang et al [5] 2015 Yes Yes No Xu et al [27] 2013 Yes Yes No Liu et al [19] 2011 Yes Yes Yes Duan et al [43] 2017 Yes Yes No Sui et al [28] 2017 [106] 2017 Yes Yes No Athira et al [107] 2018 Yes Yes No Gopinath et al [108] 2017 No Yes No Dodo et al [109] 2019 Yes Yes No Duan et al [110] 2015 Yes Yes No Lang et al [111] 2017 Yes Yes No Niu et al [112] 2014 Yes Yes No Rossant et al [113] 2015 Yes Yes No Tian et al [114] 2015 Yes Yes No Huang et al [80] 2019 No Yes Yes Nath et al [82] 2018 Yes Yes Yes Hassan and Hassan [81] 2019 Yes Yes Yes Hassan et al [1] 2016 Yes Yes Yes Fang et al [115] 2017 Yes Yes Yes the B-scans [93,94]. The OCTID was the only publicly available database found with only cases of macular holes pathology [95].…”

“…Other techniques are: multi-scale spatial pyramid (MSSP), it captures the geometry of retina at multiple scales [19]; geodesic distance method (GDM), it can locate pixels in boundaries of layers [43]; convolutional neural network (CNN), it is pooling, which is a non-linear down-sampling [28,108,120]; dynamic programming (DP), it is a method that divide problems and solves each one separately [65,100,102]; canny edge detection, it is an algorithm that detects edges [97,99,104]; markov gibbs random field (MGRF), it allows to derive a global texture description by specifying local properties of textures [98]; loosely coupled level sets (LCLS), it is a technique that uses local intensity variations to segment layers [101]; structure tensor, it utilizes the gradient of a point with neighborhood to get directions of segmentation [102,104]; Randon Forest (RF), that train the data to estimate boundary probabilities [111,121,122]; and OTSU algorithm, it is used to perform automatic image thresholding [102,107,112]. In Figure 3 is shown an example of segmentation approach applied to an OCT image, in which the top boundary of the ILM and RPE layers are highlighted.…”

Image processing methods for computer-aided screening for disease

“…Recently (H. Fu et al 2018) presented a novel ensemble network based the application of different CNNs to the global fundus image and to different versions of optic disc region. Finally, in OCT there are also recent studies applying CNNs for glaucoma detection (Muhammad et al 2017) or the segmentation of layers (Gopinath, Rangrej, and Sivaswamy 2018;Fang et al 2017). In the next table we present a summary of the datasets used and the results reported.…”