利用基於U-Net的卷積神經網路及殘差長短期記憶分割AMD患者的光學相干斷層掃描影像中的視網膜層邊界

Abstract

老年性黃斑部病變（age-related macular degeneration, AMD）是一種常見的眼科疾病，在各個國家中都會引起中心視力的逐漸退化。它的特徵是在黃斑部出現隱節（drusen），並伴有脈絡膜新生血管（choroidal neovascularization, CNV）或地圖狀萎縮（geographic atrophy, GA）。它們的大小，數量和位置可作為疾病進展的生物標記。光學相干斷層掃描（optical coherence tomography, OCT）是獲取視網膜三維影像的一種快速且無創的方法，並且越來越多地用於監測AMD的發作和進展。AMD疾病的嚴重程度很可能由隱節和地圖狀萎縮的定量確定。由於手動分割OCT影像既費時又主觀，因此有必要開發自動圖層分割演算法。本文提出並實現了一種深度學習的OCT影像自動分割方法。利用一種以U-net架構為基礎的語義分割網路加上遞歸神經網路的架構進行分割。實驗結果表明，與其他最新方法相比，該方法大大降低了錯誤率。

Age-related macular degeneration (AMD) is a common eye disease that causes a gradual deterioration of central vision in various countries. It is characterized by the appearance of drusen in the macula, accompanied by choroidal neovascularization (CNV) or geographic atrophy. Their size, number, and location can serve as biomarkers for disease progression. Optical coherence tomography (OCT) is a fast and non-invasive way of obtaining three-dimensional images of the retina and is increasingly used to monitor the onset and progression of AMD. The severity of AMD disease is likely to be determined from the quantification of drusen and geographic atrophy. However, manual segmentation of OCT images is time-consuming and subjective, it is necessary to develop an automatic layer segmentation algorithm. In this paper, we propose and implement a deep learning OCT image automatic segmentation method. Using a U-net-based fully convolutional architecture and a recursive neural network for image segmentation. Experimental results show that compared with other recent methods, this method greatly reduces the error rate.