- 學位論文
利用卷積神經網路從黃斑部光學同調斷層掃描影像分辨健康及好發黃斑前膜的亞健康眼睛
Differentiation of Normal Eyes and Sub-healthy Ones Susceptible to Epiretinal Membrane by Macular Optical Coherence Tomography Using Convolutional Neural Networks
摘要
過去普遍認為視網膜黃斑部的結構差異只是個體上的差別，不具有臨床上的意義，但隨著近年在研究黃斑部具有寬中心凹（wide-based）的眼睛時發現這些具有寬中心凹的眼睛有近83%是女性，這些具有寬中心凹的眼睛雖然其視力與功能皆完全正常，但對側眼卻有大約52%的比例產生黃斑前膜（epiretinal membrane; ERM）。另一個研究在觀察單側黃斑前膜的健康對側眼時發現，發現這些健康對側眼有著寬中心凹，以及相較於正常人有著較薄的中心凹厚度（central foveolar thickness; CFT）、較薄的中心視網膜厚度（central subfield thickness; CST）以及更大的中心凹無血管區（foveal avascular zone; FAZ），而這些健康眼睛的視力與功能也完全正常。另外也有研究利用卷積神經網路分析視網膜黃斑部的光學同調斷層掃描（optical coherence tomography; OCT）可以進行年齡的預測，甚至可以藉由觀察中心凹的輪廓預測性別。這些可能都代表著看似只是個體差異的視網膜黃斑部結構特徵，其實包含著許多資訊與臨床上的意義。 已廣泛應用在影像分類或者影像分割的卷積神經網路（convolution neural network; CNN）近年來也開始用於醫學影像，除了CNN在影像處理上有著強大的性能，CNN具有自動找出影像特徵的能力，並且透過可視化工具呈現特徵，以利分析；而OCT能夠以非侵入式的方式獲得高解析度的樣本影像。本論文結合深度學習以及OCT影像，使用卷積神經路分辨4866組來自2288位病人的健康、黃斑前膜與其他黃斑部結構異常（不包含黃斑前膜）之視網膜黃斑部OCT影像，CNN模型在分辨影像屬於健康、黃斑前膜或者其他黃斑部結構異常（不包含黃斑前膜）時可以取得80%以上的分辨準確率。除此之外，也利用經過OCT影像預訓練之CNN模型分辨1155組810為病患的健康眼睛與單側患有黃斑前膜的健康對側眼睛之OCT影像，再將資料依照性別分類後，CNN模型可以取得66.6±6.1％的分辨健康眼睛與單側患有黃斑前膜的健康對側眼睛的分辨準確率，並根據可視化工具分析特徵位置。希望透過深度學習以及高解析度之OCT影像，除了幫助診斷視網膜黃斑部之病變外，也希望藉由分辨單側患有黃斑前膜的健康對側眼睛，預測未來容易產生黃斑前膜病變之眼睛，以利提早追蹤治療。
並列摘要
In the past, structural differences in the macula of the retina were generally regarded as individual variations without clinical significance. However, a recent study has discovered eyes with a wide-based macula that nearly 83% of these eyes with a wide-based macula belong to females. Although these eyes with a wide-based macula exhibit normal visual acuity and functionality, approximately 52% of the contralateral eyes develop epiretinal membranes. Another study observed the contralateral eyes of individuals with unilateral epiretinal membranes and found that, despite their normal visual acuity and functionality, they also had a wide-based macula, thinner central foveolar thickness (CFT), thinner central subfield thickness (CST), and a larger foveal avascular zone (FAZ) compared to normal individuals. These eyes also demonstrated normal visual acuity and functionality. Additionally, there have been studies utilizing convolutional neural networks (CNNs) to analyze optical coherence tomography (OCT) scans of the macula, which can predict age and even infer gender by observing the contour of the central fovea. These findings suggest that seemingly individual variations in the macular structure encompass valuable information and clinical significance. In recent years, convolutional neural networks (CNNs), widely employed in image classification and segmentation tasks, have also started to be utilized in medical imaging. Apart from their powerful performance in image processing, CNNs possess the ability to automatically extract image features and visualize them using tools for analysis. On the other hand, OCT enables the acquisition of high-resolution sample images in a non-invasive manner. This study combines deep learning with OCT images to classify OCT images of the macula from 4,866 sets of 2,288 patients into healthy, epiretinal membrane, or other macular structural abnormalities. The CNN model achieves an accuracy of over 80% in distinguishing between healthy, epiretinal membrane, and other macular structural abnormalities. Furthermore, using a pre-trained CNN model on OCT images, 1,155 sets of OCT images from 810 patients with healthy eyes and contralateral eyes with unilateral epiretinal membranes are classified after gender stratification. The CNN model achieves a discrimination accuracy of 66.6±6.1% in distinguishing between healthy eyes and contralateral eyes with unilateral epiretinal membranes, considering gender classification, and analyzing feature positions using visualization tools. The objective of this study is to utilize deep learning and high-resolution OCT images not only to aid in diagnosing macular pathologies but also to predict future eyes susceptible to epiretinal membrane development by distinguishing healthy contralateral eyes with unilateral epiretinal membranes, enabling early monitoring and treatment.