Choroidal layer segmentation in OCT images by a boundary enhancement network

Abstract

Morphological changes of the choroid have been proved to be associated with the occurrence and pathological mechanism of many ophthalmic diseases. Optical Coherence Tomography (OCT) is a non-invasive technique for imaging of ocular biological tissues, that can reveal the structure of the retinal and choroidal layers in micron-scale resolution. However, unlike the retinal layer, the interface between the choroidal layer and the sclera is ambiguous in OCT, which makes it difficult for ophthalmologists to identify with certainty. In this paper, we propose a novel boundary-enhanced encoder-decoder architecture for choroid segmentation in retinal OCT images, in which a Boundary Enhancement Module (BEM) forms the backbone of each encoder-decoder layer. The BEM consists of three parallel branches: 1) a Feature Extraction Branch (FEB) to obtain feature maps with different receptive fields; 2) a Channel Enhancement Branch (CEB) to extract the boundary information of different channels; and 3) a Boundary Activation Branch (BAB) to enhance the boundary information via a novel activation function. In addition, in order to incorporate expert knowledge into the segmentation network, soft key point maps are generated on the choroidal boundary, and are combined with the predicted images to facilitate precise choroidal boundary segmentation. In order to validate the effectiveness and superiority of the proposed method, both qualitative and quantitative evaluations are employed on three retinal OCT datasets for choroid segmentation. The experimental results demonstrate that the proposed method yields better choroid segmentation performance than other deep learning approaches. Moreover, both 2D and 3D features are extracted for statistical analysis from normal and highly myopic subjects based on the choroid segmentation results, which is helpful in revealing the pathology of high myopia. Code is available at https://github.com/iMED-Lab/Choroid-segmentation.

Keywords: boundary segmentation; choroidal layer; deep learning; high myopia; optical coherence tomography.

FIGURE 1

Examples of a 3D OCT volume and 2D OCT B-scan image. The choroid-sclera interface (indicated by the red dashed line) is ambiguous and difficult to extract compared to the other boundaries.

FIGURE 2

An overview of the proposed boundary enhancement framework for choroid segmentation. A novel BEM is incorporated into each encoder/decoder layer of the proposed framework. In addition, a pre-trained VGG network is utilized to calculate the specific boundary perceptual loss to improve the choroidal boundary segmentation with the guidance of the soft point map generated from ground truth.

FIGURE 3

Two different types of point maps extracted from the same OCT B-scan. The original points map was generated using binary disk as (Lee et al., 2020), while the soft map was generated based on a two-dimensional Gauss function. All points were extracted from the same boundary of ground truth.

FIGURE 4

The architecture of the proposed BEM (A). The BEM consists of three parallel branches including FEB (B), CEB (C) and BAB (D), which achieve enhancement of boundary information from the feature, channel and spatial perspective, respectively.

FIGURE 5

Examples of the original OCT images from (A) COSTA-H dataset, (B) COSTA-T dataset, (C) COSTA-B dataset (6 bit-depth), and (D) COSTA-B dataset (12 bit-depth).

FIGURE 6

The visualization of the example result of choroid segmentation on the COSTA-H dataset. The first image is the original image, the second image is the ground truth, and the next few images are the results of different methods of segmentation: the specific methods are marked in the upper left corner of the image. White denotes a correctly segmented choroidal area, red denotes over-segmentation, and blue denotes under -segmentation.

FIGURE 7

Results of different choroid segmentation methods in boundary detection. The name of the method is shown at the upper left corner of each image.

FIGURE 8

Comparison results of other choroid segmentation methods in boundary detection. The specific methods are marked in the upper left corner of the image.

FIGURE 9

Trend of Dice and IoU results of different segmentation methods in different bit depth images. (A), (C) are the result on the COSTA-H dataset, and (B), (D) are the result on the COSTA-T dataset.

FIGURE 10

Ablation results on COSTA-H and COSTA-T datasets. The blue bars denote the quantitative results of the baseline network U-Net. The orange bars denote the segmentation results of the network with FEB. The gray bars denote the segmentation results of the network with FEB and CEB. The yellow bars denote the segmentation results of the network with FEB, CEB, and BAB. The red bars denote the segmentation results of the network with FEB, CEB, BAB, and BP-Loss.

FIGURE 11

Choroidal layer thickness map in normal and highly myopic subjects using Early Treatment Diabetic Retinopathy Study (ETDRS) circles of 1 mm, 3 mm, and 6 mm. The standard ETDRS subfields dividing the macula into 9 subfields. CFT: Central foveal thickness; TIM: Temporal inner macula; NIM: Nasal inner macula; SIM: Superior inner macula; IIM: Inferior inner macula; TOM: Temporal outer macula; NOM: Nasal outer macula; SOM: Superior outer macula; IOM: Inferior outer macula. (A) denotes the thickness map in normal subject, (B) denotes the 9 subfields of macula, (C,D) denote the average choroidal thickness [μm] of subfields in normal subjects and highly myopic, respectively.