School of Engineering Department of Computer Science and Engineering 99 Deep Learning for Ophthalmology Image Analysis Supervisor: CHEN Hao / CSE Student: WANG Yueying / DSCT Course: UROP1100, Fall UROP2100, Spring Traditionally, Diabetic macular ischaemia (DMI) has been defined and graded on the angiographic basis of the enlarged and irregular avascular area of the fovea. Using advanced optical coherence tomography angiography (OCTA) to identify early DMI phenotypes is the key to successful treatment of DMI. However, due to the patient-related artifacts, technically related artifacts or both related artifacts, the judgment result of deep learning models is not reasonable. In this undergraduate research opportunity project, we tried to add prior knowledge to the model according to auxiliary learning and multi-task learning, and relabeling the dataset to achieve better results. Deep Learning for Ophthalmology Image Analysis Supervisor: CHEN Hao / CSE Student: ZHANG Weiwen / DSCT Course: UROP2100, Fall Optical coherence tomographic angiography (OCTA) is a novel non-invasive technology that can capture retinal and choroidal microvascular images without any invasive operation. The scan region of OCTA is typically 3×3-mm and 6×6-mm. Although 6×6-mm has a larger field of view, its defect is the relatively low resolution compared with 3×3-mm due to under-sampling. This paper first proposed an unsupervised highresolution reconstruction pipeline in the frequency domain based on the structure of CycleGAN and made progress comparing it with vanilla CycleGAN. Based on it, this report continued to propose a network that does not need interpolation algorithms as preprocessing and constrains the training in the frequency domain. This paper has been published in top-tier conference MICCAI: https://link.springer.com/chapter/10.1007/978-3-031-16434-7_62
RkJQdWJsaXNoZXIy NDk5Njg=