基於深度學習之青光眼早期診斷:結合黃斑部光學同調斷層掃描厚度圖譜與視野資訊

Abstract

根據世界衛生組織統計，青光眼為全球致盲率第二高的疾病。隨著全球人口老化，青光眼的罹患人數也越來越多。青光眼是一種慢性且不可逆的視神經病變，其特徵為視網膜神經節細胞的凋亡以及視野功能的損失，目前對於青光眼有效的解決策略為早期診斷、早期治療，然而青光眼的早期診斷是一件極具挑戰性的任務，在青光眼早期通常沒有明顯症狀，臨床醫師須綜合多種檢查進行診斷，但根據臨床醫師經驗的不同，在判斷上仍會有差異且較為主觀，因此青光眼電腦輔助診斷系統可以幫助臨床醫師提供客觀的建議，以輔助臨床醫師進行決策。 目前青光眼之電腦輔助診斷的相關研究主要以視野檢查、眼底鏡影像以及Optical coherence tomography (OCT)為主，然而視野資訊與眼底鏡影像在青光眼的早期診斷上的靈敏度仍然不足。OCT能提供較精細的視網膜結構性資訊，但大多是以各家廠商所計算出來的厚度資訊，再以機器學習的方式進行訓練，而這些厚度資訊皆是區域性的平均數值，因此有一些局部的資訊可能會被忽略。另外，目前OCT在深度學習應用上皆單一探討結構資訊，未將具功能性的視野資訊納入考量，不符合目前臨床醫師之診斷方式。在資料集方面，目前多數研究並沒有考慮到青光眼的嚴重程度，而在對照組的樣本皆是採用健康樣本，並不符合目前實際臨床上的需求。 因此本研究將開發一基於深度學習的早期青光眼診斷系統對早期青光眼與非青光眼進行分類，青光眼樣本透過Glaucoma Staging System進行嚴重程度的分期，非青光眼樣本包括了健康樣本以及青光眼高風險族群樣本，此系統分為分割與分類兩部分，分割的部分透過本研究開發的GCC layer segmentation net對黃斑部Ganglion cell complex (GCC)層進行分割，接著再透過Multiple-view GCC layer segmentation net進行微調，再將分割好的GCC層使用最近距離法計算出厚度圖。分類的部分將厚度圖以及差異圖利用本研究開發的Attention Convolution neural network進行分類，再將兩種不同的視野檢查資訊(Standard Automated Perimetry- Mean Defect、Standard Automated Perimetry- square root of loss variance、Pulsar Perimetry - Mean Defect、Pulsar Perimetry - square root of loss variance)與Attention Convolution neural network進行結合。 從分割結果顯示，本研究之GCC layer Segmentation net能夠有效的對黃斑部GCC層進行分割，其Dice coefficient青光眼為0.9754±0.0008、非青光眼為0.9822±0.0014，其Hausdorff distance青光眼為10.59±0.67 μm、非青光眼為8.47±0.72 μm，透過Multiple-view segmentation對邊緣進行修正後，其Dice coefficient青光眼為0.9776±0.0104，非青光眼為0.9846±0.0060，其 Hausdorff distance青光眼為9.26±0.66 μm，非青光眼為7.21±0.68 μm，從結果圖上來看，可以將崎嶇的邊緣進行平滑化，補足2D分割模型對3D影像分割所造成的上下文不連續問題。 從分類結果顯示，本研究之Attention Convolution neural network透過深度學習對厚度圖進行特徵提取並加上差異圖強化視網膜厚度圖缺損的空間位置，結果顯示早期青光眼AUC為0.8488±0.0034，中晚期青光眼AUC為0.8715±0.0049，不分期青光眼AUC為0.8520±0.0031。本研究也單獨探討兩種不同類型之視野資訊對於青光眼之分類結果，結果顯示早期青光眼AUC為0.8073±0.0095，中晚期青光眼AUC為0.9997±0.0006，不分期青光眼AUC為0.8384±0.0037。最後將兩種不同的視野檢查資訊與ACNN進行結合，結果顯示早期青光眼AUC為0.8931±0.0023，中晚期青光眼AUC為0.9947±0.0027，不分期青光眼AUC為0.9075±0.0021，證明了兩種不同的視野資訊與厚度圖資訊的結合有助於提升早期青光眼的診斷能力，並且結果高於單一的厚度資訊以及視野資訊的結果。

According to the World Health Organization, glaucoma is the second leading cause of blindness in the world. As the population ages growing, the number of people suffering from glaucoma is increasing. Glaucoma is an irreversible and progressive optic neuropathy characterized by the apoptosis of retinal ganglion cells and loss of visual field. The current effective solution for glaucoma is “early diagnosis and treatment”. However, Diagnosis of early glaucoma is a challenging task, because there are no obvious symptoms in the early stage of glaucoma. Clinicians must make a diagnosis based on variety of examinations, and diagnosis by their subjective opinions. The development of computer-aided diagnosis system can provide objective opinions and make clinical decisions for clinicians. Research on computer-aided diagnosis of glaucoma is mainly based on visual field examination, fundus image and Optical coherence tomography. The sensitivity of visual field information and fundus image is still insufficient in the diagnosis of early glaucoma. Optical coherence tomography can provide more detailed retinal thickness information. At present, most of the thickness information is calculated by various manufacturers, and then trained by machine learning. However, these thickness information are regional average values. Some local information may be ignored. In addition, the current OCT research with deep learning are focus on structure information, and does not consider the functional information, which is not in line with diagnosis method of clinicians. In terms of dataset, most of the literature does not consider the severity of glaucoma, and the control samples are all healthy samples, which does not meet the current actual clinical needs. Therefore, this study will develop an early glaucoma diagnosis system. Glaucoma samples are staged through the Glaucoma Staging System. Non-glaucoma samples include healthy samples and high risk glaucoma samples. This system is divided into segmentation part and classification part. In segmentation part, this study develops a GCC layer segmentation net (GCCLS net) to segment the GCC layer of the macula, and then fine-tunes with Multiple-view GCC layer segmentation net (Multiple-view GCCLS net). The segmented GCC layer uses the nearest distance method to calculate the thickness map. In the classification part, this study develops the Attention Convolution neural network with the thickness map and the difference map, and then add two different perimetry information (Standard Automated Perimetry-Mean Defect, Standard Automated Perimetry-square root of loss variance, Pulsar Perimetry-Mean Defect, Pulsar Perimetry-square root of loss variance) to classify glaucoma and non-glaucoma. The segmentation results show that the GCCLS net of can effectively segment the GCC layer of the macula. Its dice coefficient is 0.9754±0.0008 for glaucoma, 0.9822±0.0014 for non-glaucoma, and its Hausdorff distance for glaucoma is 10.59±0.67 μm. Glaucoma is 8.47±0.72 μm. Through the fine-tuning with Multiple-view GCCLS net, its dice coefficient is 0.9776±0.0104 for glaucoma and 0.9846±0.0060 for non-glaucoma. Its Hausdorff distance is 9.26±0.66 μm for glaucoma and 7.21±0.68 μm for non-glaucoma. From the result picture, the rugged edges can be refined to make up for context discontinuities caused by the 2D segmentation model on the 3D image. The classification results show that the Attention Convolution neural network uses deep learning to extract the features of the thickness map and enhance the spatial location of the retinal defect by the difference map. The AUC of early glaucoma is 0.8488±0.0034, and the AUC of moderate to end stage glaucoma is 0.8715±0.0049, the AUC of non-stage glaucoma was 0.8520±0.0031. This study also separately discussed the classification results of two different types of perimetry information for glaucoma. The results showed that the AUC of early glaucoma was 0.8073±0.0095, the AUC of moderate to end stage glaucoma was 0.9997±0.0006, and the AUC of non-staged glaucoma was 0.8384±0.0037. Finally, the two different perimetry information were combined with ACNN. The results showed that the AUC of early glaucoma was 0.8931±0.0023, the AUC of moderate to end stage glaucoma was 0.9997±0.0006, and the AUC of non-staged glaucoma was 0.9075±0.0021. It’s proved that the combination of perimetry information and thickness map information can helps improve the diagnostic ability of early glaucoma, and is higher than the result of single thickness information and visual field information.