School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China.
Department of Ophthalmology, The First Affiliated Hospital with Nanjing Medical University, Nanjing, China.
Med Phys. 2021 Dec;48(12):7773-7789. doi: 10.1002/mp.15315. Epub 2021 Nov 10.
To robustly segment retinal layers that are affected by complex variety of retinal diseases for optical coherence tomography angiography (OCTA) en face projection generation.
In this paper, we propose a robust retinal layer segmentation model to reduce the impact of multifarious abnormalities on model performance. OCTA vascular distribution that is regarded as the supplements of spectral domain optical coherence tomography (SD-OCT) structural information is introduced to improve the robustness of layer region encoding. To further reduce the sensitivity of region encoding to retinal abnormalities, we propose a multitask layer-wise refinement (MLR) module that can refine the initial layer region segmentation results layer-by-layer. Finally, we design a region-to-surface transformation (RtST) module without additional training parameters to convert the encoding layer regions to their corresponding layer surfaces. This transformation from layer regions to layer surfaces can remove the inaccurate segmentation regions, and the layer surfaces are easier to be used to protect the retinal layer natures than layer regions.
Experimental data includes 273 eyes, where 95 eyes are normal and 178 eyes contain complex retinal diseases, including age-related macular degeneration (AMD), diabetic retinopathy (DR), central serous chorioretinopathy (CSC), choroidal neovascularization (CNV), and so forth. The dice similarity coefficient (DSC: %) of superficial, deep and outer retina achieves 98.92, 97.48, and 98.87 on normal eyes and 98.35, 95.33, and 98.17 on abnormal eyes. Compared with other commonly used layer segmentation models, our model achieves the state-of-the-art layer segmentation performance.
The final results prove that our proposed model obtains outstanding performance and has enough ability to resist retinal abnormalities. Besides, OCTA modality is helpful for retinal layer segmentation.
为了稳健地分割受多种视网膜疾病影响的视网膜层，以便生成光学相干断层扫描血管造影（OCTA）的面投影。
在本文中，我们提出了一种稳健的视网膜层分割模型，以减少多种异常对模型性能的影响。将 OCTA 血管分布作为光谱域光相干断层扫描（SD-OCT）结构信息的补充，引入到层区域编码的稳健性中。为了进一步降低区域编码对视网膜异常的敏感性，我们提出了一个多层次的细化（MLR）模块，可以逐层细化初始层区域分割结果。最后，我们设计了一个无需额外训练参数的区域到表面转换（RtST）模块，将编码层区域转换为相应的层表面。这种从层区域到层表面的转换可以去除不准确的分割区域，并且层表面比层区域更容易用于保护视网膜层的性质。
实验数据包括 273 只眼睛，其中 95 只眼睛正常，178 只眼睛包含复杂的视网膜疾病，包括年龄相关性黄斑变性（AMD）、糖尿病性视网膜病变（DR）、中心性浆液性脉络膜视网膜病变（CSC）、脉络膜新生血管（CNV）等。在正常眼中，浅层、深层和外层视网膜的 DSC（%）分别达到 98.92、97.48 和 98.87，在异常眼中分别达到 98.35、95.33 和 98.17。与其他常用的层分割模型相比，我们的模型实现了最先进的层分割性能。
最终结果证明，我们提出的模型具有出色的性能，并且有足够的能力抵抗视网膜异常。此外，OCTA 模态有助于视网膜层分割。
