利用卷積神經網路對黃斑部病變的視力進行預測之研究

Abstract

黃斑部皺褶，是一種慢性眼疾，經常發生在年長者身上，患者視網膜的黃斑 部會產生皺摺，進而影響視力。不過，雖然已知此疾病對於視力有非常重大的影 響，但在同樣患有此疾病的患者當中，卻可能擁有不同的視力分布，有些病人的 視力可能僅僅只有 0.1，有些病人卻能夠擁有高達 1.0 的視力。視力的差異難以單 純地依靠肉眼檢視醫學影像來判斷，因此，以深度學習為基礎的電腦視覺將可能 是一個有效之方法。深度學習在這幾年來可以說是蓬勃發展，尤其是在影像辨識方面更是有著相 當優異的表現，本論文將使用 Resnet18、Resnet50、MobilenetV2、ShuffleV2 這四 種神經網路來加以分析，透過卷積神經網路強大的圖形識別能力，來幫助我們找 到在患有黃斑部皺褶的病人的黃斑部之中影響視力最為關鍵的部分。本論文所使 用的資料集是採用台大醫院眼科所提供的 angio retina 影像，它是一種使用了光學 原理成像的眼底血管影像，由於本論文中所使用到的資料集較難以蒐集，所以在 數量上比較稀少，因此除了針對資料集做了資料增強來增加資料集的數量外，另 外還有使用投票法、K 折交叉驗證等方法，來提升模型的表現，在實驗的最後， 本論文採用了 Grad-CAM++這個工具，使訓練結果可以視覺化，以熱像圖的方式 描繪出卷積神經網路所關注的區域，希望此有助於眼科醫師的臨床判斷。

Epiretinal Membrane (ERM) is a chronic eye disease that often occurs in the elderly. The macular area of the retina of the patient will be wrinkled, which will affect vision. However, although this disease is known to have a very significant impact on vision, the patients with this disease may have different visual acuity. Some patients’ visual acuity (VA) may only be 0.1, while some patients’ VA can be 1.0. It is difficult to judge the difference in vision through medical images by naked eyes. Therefore, computer vision based on deep learning may be an effective method.Deep learning has flourished in recent years, especially in image recognition. In this thesis, we will use Resnet18, Resnet50, MobilenetV2, and ShuffleV2 these four neural network models to help us to find the most critical part of the macula of patients with ERM. The dataset used in this thesis is “angio retina”, which is provided by the Department of Ophthalmology of National Taiwan University Hospital. It is a blood vessel image of the eye. Since it is difficult to collect the images, the amount of the images is relatively small. Thus, we use data augmentation to increase the amount of images. In addition, we used voting and the K-fold cross-validation to improve the performance of the model. At the end of the experiment, we used Grad-CAM++ to visualize the training results. It is expected that the experimental results can really help ophthalmologists clinically.