Screening and Segmentation of Polypoidal Choroidal Vasculopathy on FA Images Using U-Net

• Main Authors: LEE, I-YANG, 李亦洋
• Other Authors: LIN, WEI-YANG
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/uum8n7

碩士 === 國立中正大學 === 資訊工程研究所 === 107 === With the advancement of medical technology, the elderly population has grown rapidly. Age-related Macular Degeneration (AMD) is a common eye disease in the elderly people. Among the subtypes of AMD, Polypoidal Choroidal Vasculopathy (PCV) appears to be predominant in Asian populations. It can quickly affect the patient’s vision and cause permanent damage. The clinical diagnosis of PCV typically requires two kinds of angiography, Fluorescein Angiography (FA) and IndoCyanine Green Angiography (ICGA). In this thesis, we develop a novel approach to perform early screening and segmentation of PCV using only FA images. This is a challenging task because the diagnosis of PCV requires another imaging modality (i.e., ICGA) in the current clinical practice. We hope that our proposed method could speed up the diagnosis process of PCV. In our proposed framework, we use convolutional neural network to extract features from the FA images. The extracted features are forwarded to the SVM classifier for early screening of PCV. Furthermore, we modify the U-Net architecture and use it to segment PCV lesion in FA images. To validate our proposed method, we have conducted several experiments using the EVEREST dataset and two other clinical datasets. For the screening of PCV, our proposed method achieves accuracy of 92.78\% and 89.13\% on the two clinical datasets. For the segmentation of PCV, our proposed method achieves segmentation accuracy of 77.60\% on the EVEREST dataset. Comparison with previous studies are also reported in the thesis.