息肉狀脈絡膜血管病變 (Polypoidal Choroidal Vasculopathy, PCV)是好發於亞洲人身上的一種眼科疾病,嚴重會導致失明。 PCV 是屬於AMD (Age-Related Macular Degeneration,老年性黃斑部病變)的一種類型,目前診斷用的方式有 FA (Fluorescein Angiography,眼底螢光血管攝影)及 ICGA (IndoCyanine Green Angiography,眼底循血綠血管攝影),通常一開始醫生會以 FA 檢查病人是否患有 AMD,但FA 由於穿透力不足因難以發現位於患部位於較深層的 PCV,因此臨床診斷上醫生懷疑病人有 PCV時會施以 ICG檢查 我們的系統使用卷積神經網路進行 PCV偵測,其最主要的優點在於不需設計影像特徵,而是通過大量的資料由電腦自動學習具代表性的特徵,我們的主要目標是以 ICG 影像訓練我們的神經網路架構並偵測 PCV,次要目標則是於 FA 影像上偵測 PCV,由於 FA影像上有大量的螢光劑滲漏使偵測病變區域的難度極高,因此在本論文中只進行嘗試性的實驗。我們採用EVEREST的資料庫進行實驗,並且系統在PCV病變中的其中一個病徵(polyps)上取得比現有方法更好的偵測準確度。
Polypoidal choroidal vasculopathy (PCV) is a subtype of age-related macular degeneration (AMD) and is prevalent from 40$\%$ to 50$\%$ among the Asian patients with exudative maculopathy. Currently, indocyanine green (ICG) angiography is the most widely adopted imaging modality for the diagnosis of PCV. In this thesis, we propose a novel method to detect PCV lesion in an ICG sequence using convolution neural network (CNN). In particular, we construct an eight-layer CNN model and use it to determine whether a pixel location belongs to PCV or not. We have conducted experiments using ICG sequences from the EVERSET dataset. We carry out leave-one-out cross validation to evaluate the performance of our proposed method. The results show that our proposed method achieves comparable performance as the state-of-the-art method.