Improving OCT Image Segmentation of Retinal Layers by Utilizing a Machine Learning Based Multistage System of Stacked Multiscale Encoders and Decoders

Abstract

Optical coherence tomography (OCT)-based retinal imagery is often utilized to determine influential factors in patient progression and treatment, for which the retinal layers of the human eye are investigated to assess a patient's health status and eyesight. In this contribution, we propose a machine learning (ML)-based multistage system of stacked multiscale encoders and decoders for the image segmentation of OCT imagery of the retinal layers to enable the following evaluation regarding the physiological and pathological states. Our proposed system's results highlight its benefits compared to currently investigated approaches by combining commonly deployed methods from deep learning (DL) while utilizing deep neural networks (DNN). We conclude that by stacking multiple multiscale encoders and decoders, improved scores for the image segmentation task can be achieved. Our retinal-layer-based segmentation results in a final segmentation performance of up to 82.25±0.74% for the Sørensen-Dice coefficient, outperforming the current best single-stage model by 1.55% with a score of 80.70±0.20%, given the evaluated peripapillary OCT data set. Additionally, we provide results on the data sets Duke SD-OCT, Heidelberg, and UMN to illustrate our model's performance on especially noisy data sets.

Keywords: OCT biomarkers; OCT segmentation; computer vision and pattern recognition; deep learning; machine learning; ophthalmology; ophthalmology diseases.

Figure 1

Overview of our public data sets for qualitative assessment, Duke SD-OCT [14] (top row), UMN [54] (middle row), and Heidelberg [55] (bottom row), with and their annotations, if available. For the Duke SD-OCT data set, the retinal layers are highlighted in terms of different levels of brightness. For the UMN data set, areas with fluids are additionally available (highlighted in green). For the Heidelberg data set, no annotations are available.

Figure 2

Overview of our data set with two selected samples and their visualized retinal layers following the color scheme presented in Table 2 as a legend for the present visualizations. Table 2 gives an overview of our data set with its samples and the proposed data set split ratio. In the top row, exemplary peripapillary OCT images are shown with their annotations in the bottom row.

Figure 3

The model architecture and framework of our proposed model. (a) The general model used to stack multiple encoders and decoders is illustrated. (b) The model with two stacks of encoders and decoders is shown. For visualization purposes, cyan and orange represent the input and the output, light cyan represents the first stack with encoder 1 and decoder 1, and light orange represents the second stack with encoder 2 and decoder 2.

Figure 4

Visualization of the quantitative results with (a) one stack and (b) two stacks from Table 3 and Table 4.

Figure 5

Qualitative peripapillary data set with the original, ground truth segmentation, and segmentation model results as a color-coded overlay. The individual, class-wise Dice scores are shown in Table 6.

Figure 5

Qualitative peripapillary data set with the original, ground truth segmentation, and segmentation model results as a color-coded overlay. The individual, class-wise Dice scores are shown in Table 6.

Figure 6

Qualitative results based on the Heidelberg OCT data set described in Section 2.1.4. For the Heidelberg OCT data set, no annotations are available. We consider the Heidelberg data set to be a particularly noisy data set. This is apparent when comparing Figure 5 with (c) and (k): General image noise and artifacts are observed.

Figure 7

Qualitative results based on the Duke SD-OCT data set described in Section 2.1.2. The ground truth comprises fluid and non-fluid manual annotations of eight boundaries.

Figure 8

Qualitative results based on the UMN data set described in Section 2.1.3. The ground truth comprises the retinal fluid regions.