應用於視覺敏銳度預測之彈性化插入式模組

Abstract

黃斑前膜(ERM)是一種由纖維膜覆蓋黃斑部引起的慢性眼疾，其臨床表現取決於膜的厚度及收縮能力，輕則毫無病徵，重則恐致中心視覺永久性喪失。手術是ERM唯一的治療方式，目前國際尚無制定最佳手術時機的標準，一般以患者術前視覺敏銳度(俗稱視力)受損的程度，評估其術後恢復的成效。有鑑於視網膜影像種類繁多、結構複雜多變、病灶隱蔽不易察覺，單憑我們肉眼觀察難以立即區辨視力的細微變化。近年來，深度學習在視網膜疾病檢測中的應用日趨成熟，但在疾病檢測的子類別中，與視力預測相關的研究卻寥寥無幾。因此，我們提出一個易於應用在各式骨幹網路的插入式模組，它結合了背景分割和特徵融合技術，有效整合骨幹網路每一區塊的輸出特徵。在僅有影像層級註釋的情況下，利用光學同調斷層掃描(OCT)、光學同調斷層掃描血管造影(OCTA)影像及視網膜厚度(RT)圖來預測ERM患者的視力。實驗結果表明OCT、OCTA影像及RT圖，其整體十折交叉驗證的平均準確度分別達80.558% (+8.176%)、81.502% (+5.502%)、82.534% (+7.112%)。隨後便利用Grad-CAM進行特徵視覺化，以輔助醫師及時診斷並採取適當措施。

Epiretinal membrane (ERM) is a chronic eye disease caused by the fibrous membrane covering the macula. Its clinical manifestations depend on the thickness and contractility of the membrane, ranging from asymptomatic to permanent loss of central vision. Surgery is the only treatment for ERM. Currently, there is no international standard for the optimal timing of surgery. Generally, the extent of preoperative visual impairment of the patient is used to evaluate the effect of postoperative recovery. In view of the wide variety of retinal images, the complex and changeable structure, and the hidden and difficult to detect lesions, it is hard to immediately distinguish subtle changes in visual acuity with naked eye alone. In recent years, the application of deep learning in retinal disease detection has become more and more mature, but in the subcategory of disease detection, there are few studies related to visual acuity prediction.Therefore, we propose a plug-in module that is easy to apply to various backbone networks. It combines background segmentation and feature fusion techniques to effectively integrate the output features of each block of the backbone network. In the case of only image-level annotation, optical coherence tomography (OCT), optical coherence tomography angiography (OCTA) images, and retinal thickness (RT) maps were used to predict visual acuity in patients with ERM. The experimental results show that the mean accuracy of the overall 10-fold cross validation of OCT, OCTA images and RT Maps is 80.558% (+8.176%), 81.502% (+5.502%), and 82.534% (+7.112%). Then, Grad-CAM is used for feature visualization to assist physicians in making timely diagnosis and taking appropriate measures.