深度學習模型於眼底鏡影像篩檢羥氯喹視網膜病變之應用

Abstract

視網膜病變可由多種原因引起，嚴重影響患者視力和生活品質。傳統上，視網膜黃斑部病變主要透過光學相干斷層掃描進行檢測，但由於其斷層掃描的特性，光學相干斷層掃描只能檢測黃斑部等視網膜中央區域，限制了其能檢測的視網膜疾病種類。眼底攝影是另外一種檢測視網膜病變的方式，其能夠檢測更廣泛範圍的視網膜，但其影像特性導致眼科醫師難以用其診斷黃斑部病變。本研究著重在以眼底鏡影像檢測自體免疫性疾病用藥羥氯喹導致的視網膜黃斑部病變，是第一篇嘗試這麼做的研究。我們所提出的基於深度學習的電腦輔助診斷系統，包括影像前處理、黃斑部遮罩生成與視網膜病變預測。在影像前處理，會先調整圖片大小並進行裁切。同時，影像會被輸入中央凹定位模型來預測黃斑部的位置，並根據預測結果來生成黃斑部遮罩。接著，眼底鏡影像與黃斑部遮罩會被輸入分類器來判斷黃斑部是否出現病變。中央凹定位模型與分類器兩者都是以ConvNeXt網路為基底。最後，我們在訓練模型時加入了課程學習、集成學習與混合專家模型來嘗試改善模型預測的準確性和穩定性。根據實驗結果，本研究提出的電腦輔助診斷系統達到了77.73%的準確度、77.14%的靈敏度、77.86%的特異度與0.8442的ROC曲線下面積。這些結果表明使用眼底鏡影像檢測黃斑部病變具有可行性及有效性，有助於輔助眼科醫生進行診斷。

Hydroxychloroquine (HCQ), besides its antimalarial use, is also prescribed for autoimmune diseases such as rheumatoid arthritis. However, when taken long-term for chronic conditions, HCQ can exert toxic effects on the retinal macula. Consequently, patients receiving prolonged HCQ therapy must undergo regular screening for maculopathy. Maculopathy is typically detected via optical coherence tomography (OCT), but OCT’s tomographic nature limits it to imaging central retinal regions such as the macula, restricting the spectrum of detectable diseases. In contrast, color fundus photography (CFP) captures a wider retinal field and can reveal pathologies beyond OCT’s reach. Yet CFP’s image characteristics fail to display obvious visual changes, making it difficult even for expert ophthalmologists to diagnose HCQ-induced maculopathy from fundus photos alone. Therefore, this study seeks to detect HCQ-induced maculopathy using CFP images, and we believe it is the first effort in this direction. Our proposed computer aided diagnosis (CAD) system consists of three stages: image preprocessing, macular mask generation, and retinopathy classification. In preprocessing, fundus images are resized and cropped. Simultaneously, they are fed into a fovea localization network to predict the fovea’s position, from which a macular mask is generated. The cropped image and its corresponding mask are then input to a classifier that determines the presence of maculopathy. Both the fovea localization network and the classifier are built on the ConvNeXt architecture. During training, we incorporate Curriculum Learning, Ensemble Learning, and a Mixture of Experts framework to enhance prediction accuracy and robustness. In our experiments, the proposed CAD system achieved 77.73% accuracy, 77.14% sensitivity, 77.86% specificity, and an area under the ROC curve (AUC) of 0.8442. These results demonstrate that our method accurately identifies severe maculopathy, although detecting mild ones remains challenging.