UROP Proceedings 2020-21

School of Engineering Department of Computer Science and Engineering 113 Deep Learning in Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LAU Ying Yee Ava / DSC Course: UROP1100, Summer Osteoarthritis is a very prevalent disease, which creates a huge demand for automated and accurate assessment of osteoarthritis status. Past studies has looked into creating a reliable system, by using different bioindicators as a benchmark. These joint space area, minimum joint space width, osteophyte area, tibiofemoral angle, cartilage measurements, bone texture and bone size. However, the scoring is not accurate enough to reflect the severity of osteoarthritis and be able to be applied in real life. There are 3 potential directions to improve the scoring system: a multiple-model system with MRI and CT scans, using multiple bio indicators, and by reviewing the scoring method. Deep Learning in Medical Image Analysis Supervisor: CHEN Hao / CSE Student: OW YONG Chee Seng / CPEG Course: UROP1100, Summer Due to current advancements in technology, the applications of computer vision had expanded to the domain of biomedical science, and it seeks to automate tasks that the human visual system does with higher precision and accuracy. In this research, Semantic Segmentation is applied to locate and highlight different layers of Optical Coherence Tomography scans (OCT). For this task, I utilized a fully convolutional network model for the task, which is U-NET architecture to localize areas of OCT images. The rationale behind using this architecture is due to the ability of U-NET to localize and distinguish layers through segmentation of every pixel in the input image and funnel them to their category of classes. In this case, 6 classes were used as the model segments the 6 most obvious layers in OCT scans. Through the use of Annotated Retinal OCT images database (AROI Database) from Sestre Milosrdnice University Hospital Center, I have a total of 1105 OCT scans with both raw and masked as the initial dataset. After training the U-NET model with 50 epochs, the accuracy for train data is 0.98 with loss of 0.03, while testing on new images shown that the dice coefficient is 0.9 and the mean intersection over union score is 0.87. This shows that it considerably performs well over untested data. However, it is important to note that the training set had very similar images and masks as they are all scanned from 24 patients, and resource limitation towards the deep learning model. With this in mind, I am interested to further develop segmentation and localization of OCT images such that abnormalities (like Diabetic Macular Edema) can be detected with high accuracy and treatment can be provided.