Choroidal layer segmentation in OCT images by a boundary enhancement network

Wenjun Wu, Yan Gong, Huaying Hao, Jiong Zhang, Su Pan, Qifeng Yan, Yuhui Ma, Yitian ZhaoFrontiers in Cell and Developmental Biology 2022 · DOI

TL;DR

This paper introduces a novel deep learning architecture, the Boundary Enhancement Network (BEM), designed to improve the segmentation of the choroidal layer in OCT images. By specifically enhancing boundary features and incorporating expert knowledge through a novel loss function, the BEM achieves superior accuracy in segmenting this challenging layer, which is crucial for diagnosing various ophthalmic diseases.

What Problem Does This Paper Address?

Optical Coherence Tomography (OCT) is a powerful tool for imaging the eye's internal structures, including the retina and choroid. While retinal layers are relatively distinct, the boundary between the choroid and sclera (the Choroidal Scleral Interface or CSI) is often ambiguous in OCT images due to low contrast. This ambiguity makes accurate choroidal segmentation difficult for both automated methods and human experts, hindering the analysis of choroidal morphology, which is vital for understanding and diagnosing conditions like high myopia, glaucoma, and diabetic retinopathy. Existing deep learning methods, while improving accuracy, often struggle with this specific boundary challenge.

Methods

The core innovation of this paper is the Boundary Enhancement Module (BEM), which is integrated into a standard encoder-decoder architecture (specifically U-Net). The BEM comprises three parallel branches:

  1. Feature Extraction Branch (FEB): Uses dilated convolutions with varying dilation rates to capture features at different receptive fields, ensuring that boundary information across various scales is retained.
  2. Channel Enhancement Branch (CEB): Employs global average pooling and 1D convolutions to learn channel-wise weights, highlighting important feature channels for boundary detection.
  3. Boundary Activation Branch (BAB): Introduces a novel activation function designed to specifically amplify boundary-related signals while suppressing non-boundary signals, further refining boundary localization.

Additionally, the authors incorporate expert knowledge by generating "soft point maps" from ground truth annotations. These maps, created using a Gaussian function around key boundary points, are used in a Boundary Perceptual Loss (BP-Loss). This loss function, combined with a standard segmentation loss (Binary Cross Entropy), guides the network to learn more precise boundary features by comparing predicted and ground truth soft point maps using a pre-trained VGG network.

Datasets

The authors introduce a new dataset called COSTA (Choroidal OCT image for SegmenTAtion), comprising three subsets:

  • COSTA-H: 10 OCT volumes from healthy subjects, yielding 640 B-scans.
  • COSTA-T: 20 OCT volumes from healthy subjects, yielding 1280 B-scans.
  • COSTA-B: 199 annotated B-scans from a single normal subject, used to test robustness across different bit depths (6, 8, 10, and 12 bits).

COSTA-H and COSTA-T were used for training and testing (3:1 ratio), with 4-fold cross-validation. COSTA-B was used for robustness testing.

Key Insights and Ideas

  • Boundary Ambiguity is the Key Challenge: The paper correctly identifies the ambiguous choroid-sclera interface as the primary hurdle in accurate choroidal segmentation, a problem often overlooked by methods focusing solely on general segmentation.
  • Multi-Perspective Boundary Enhancement: The BEM's three-pronged approach (feature, channel, and spatial enhancement) is a clever way to tackle boundary detection from multiple angles, leading to more robust results.
  • Leveraging Expert Knowledge: The introduction of soft point maps and the BP-Loss is a significant contribution, demonstrating how to effectively inject domain-specific information into deep learning models for medical imaging.
  • 3D Analysis for Clinical Relevance: Moving beyond 2D segmentation to extract and analyze 3D choroidal features is crucial for understanding complex diseases like high myopia, where volumetric and structural changes are important.

Key Results

The proposed BEM-enhanced network achieved state-of-the-art performance on the COSTA datasets:

  • COSTA-H: Achieved a Dice coefficient of 97.06% and an IoU of 94.31%, with a superior Boundary Error (AUSDE) of 0.95 pixels for the BM and 3.00 pixels for the CSI.
  • COSTA-T: Achieved a Dice coefficient of 92.87% and an IoU of 86.91%.
  • Robustness: The model demonstrated excellent robustness across different image bit depths (6-12 bits), with performance degrading less than other methods on lower-quality images.
  • Clinical Application: Analysis of high myopia subjects revealed significantly thinner choroidal thickness compared to normal subjects across various macular regions, consistent with previous findings but now supported by more robust 3D analysis.

Limitations and Future Work

  • Multi-Class Segmentation: The current BEM is primarily designed for binary segmentation (choroid vs. background). The authors note that it may not be directly suitable for multi-layer segmentation (e.g., segmenting multiple retinal layers) without modifications to classify boundary points.
  • Dataset Limitations: While COSTA is a new contribution, the authors acknowledge that extending the dataset with more diverse disease types and increasing the volume of data would be beneficial. They also note that the "normal" subjects in their dataset might have had some degree of myopia, potentially affecting statistical analysis.
  • Future Directions: The authors suggest adapting the model for multi-class segmentation by adjusting weights for different boundary points and exploring adaptive point selection. They also aim to apply the model to a wider range of ophthalmic diseases.

Summary updated

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 <i>via</i> 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.
BibTeX
@inproceedings{wu2022choroidal,
  author = {Wenjun Wu and Yan Gong and Huaying Hao and Jiong Zhang and Su Pan and Qifeng Yan and Yuhui Ma and Yitian Zhao},
  title = {{Choroidal layer segmentation in OCT images by a boundary enhancement network}},
  booktitle = {Frontiers in Cell and Developmental Biology},
  year = {2022},
  doi = {10.3389/fcell.2022.1060241},
}
Figure