使用深度學習在超廣角眼底影像中進行周邊視網膜病灶診斷

Abstract

隨著科技日益進步，人們長期使用電子產品易導致近視，進而發生各種眼部及視網膜疾病。特別是周邊視網膜病變，它是一種異質、退化性病變，並且好發於近視患者。若不及時診斷及持續追蹤或治療，可能導致輕度失明，而較嚴重的患者可能會導致視網膜剝離而喪失視力。在診斷周邊視網膜病變的過程，眼科醫生需要從超廣角眼底鏡影像中查看周邊視網膜是否有異常的組織及病變區域。因此，本研究提出一個輔助診斷系統來協助眼科醫生診斷周邊視網膜病變，以縮短診斷所需的時間，減輕醫護人員的時間及人力成本。 本系統包含影像前處理、病灶切割及診斷三個部分。首先，超廣角眼底鏡影像和對應的標註影像將先經過影像前處理，裁剪並調整到指定的影像大小。接著，預處理後的影像將用於訓練本篇所提出的CADx系統，以同時學習病變的切割和診斷。本研究提出一個含有編碼器、解碼器及分類模組的端到端網路來進行影像訓練，尤其是在編碼器中，我們整合了pre-activation ResNet-50和數個Transformer層，共享編碼特徵以進行後續切割及診斷兩個任務。最後，系統會產生切割及預測的診斷結果。在本研究，我們使用總計816張影像來訓練及測試我們提出的系統，包含：視網膜剝離18張、視網膜裂孔24張、格子狀變性32張、無壓力性白化43張、周邊視網膜病變142張、及其他557張非周邊視網病變的影像。在切割的實驗中可以達到Dice係數為 0.8613、IoU值為 0.7563、及HD95值為 24.6219的結果，而在診斷的實驗中可以達到平均準確度97.57%、靈敏度93.86%、特異度99.14%及ROC曲線下面積AUC為 0.8737。實驗結果證實本研究提出的方法可以切割出病灶區域及產生客觀的診斷結果，並且與其他方法相比有較出色的結果。

With the advancement of technology, long-term using electronic products can easily lead to myopia and various eye diseases. In particular, peripheral retinal degeneration, a heterogeneous and degenerative lesion, predominates in myopic patients. If it is not diagnosed in time and followed up or treated continuously, it may lead to mild blindness, and more severe patients may cause retinal detachment and loss of vision. In diagnosing peripheral retinal degeneration, retinal specialists need to investigate whether there are abnormal tissues and lesion areas in the peripheral retina from ultra-widefield fundus image (UWFI). Therefore, this study proposed a computer-aided diagnosis (CADx) system to assist retinal specialists in diagnosing peripheral retinal degeneration to shorten the time required for diagnosis and reduce retinal specialists' time and labor costs. The proposed system includes image preprocessing, lesion segmentation, and diagnosis. First, the UWFI and the corresponding ground truth image will be cropped and resized to the appropriate image size. Then, the pre-processed images are used to train the proposed CAD system to simultaneously learn the segmentation and diagnosis of lesions. This study proposes an end-to-end network including an encoder, decoder, and classification block; especially in the encoder, we integrate pre-activation ResNet-50 and Transformer layers to share the encoding features for segmentation and diagnosis tasks. Finally, the system output the segmentation and the predicted diagnostic results. In this study, we used 816 images to train and test our proposed method, including 18 retinal detachments, 24 retinal breaks, 32 lattice degenerations, 43 white without pressures, 142 peripheral retinal degenerations, and 557 non-peripheral retinal images. The result of the segmentation experiment was Dice score coefficient (DSC) of 0.8613, Intersection over Union (IoU) of 0.7563, and Hausdorff distance (HD95) of 24.6219. In addition, in the diagnosis experiment, the average accuracy achieved 97.57%, sensitivity 93.86%, specificity 99.14%, and the area under the ROC curve (AUC) 0.8737. The experimental results proved that the method proposed in this study could segment the lesion area and provide objective diagnostic results that have better results than other methods.