UROP Proceedings 2020-21

School of Engineering Department of Electronic and Computer Engineering 155 Learning with Limited Anotated Data for Medical Image Diagnosis Supervisor: LI Xiaomeng / ECE Student: GANESH Vignesh / SENG Course: UROP1000, Summer Medical image analysis has come leaps and bounds since the integration of technology into the field of medical science. Machine learning and artificial intelligence have substantially improved the efficiency of medical image diagnosis by aiding medical professionals in the detection of numerous diseases. These complex models are trained rigorously to classify images based on certain parameters and can identify minute differences between images that are indistinguishable to the human eye. Deep learning models require the datasets to be labelled based on certain classifications before it is passed through for the training and testing phase. Limited annotated data has made it difficult for researchers to carry out accurate medical image analysis in a field where the margins for error are infinitesimal. This project aims to implement and improvise AI algorithms that can be used for medical image analysis with limited annotated data. Through variants of supervised learning, the main goal is to make the algorithm more intelligent by training with unlabelled data or data with few labels, thereby improving its accuracy and effectiveness. Learning with Limited Anotated Data for Medical Image Diagnosis Supervisor: LI Xiaomeng / ECE Student: GUO Jiarong / ELEC Course: UROP1100, Summer The world faces considerable challenges in terms of eye care, according to WHO, the number of visually impaired people worldwide is more than 2.2 billion. Early intervention to detect and diagnose vision could reduce the impact on at least one billion people. Based on the fundus images in RFMiD, which were captured by a retinal specialist himself at an Eye Clinic and public screening camp, we were able to build up a model based on ResNet50, which can detect the disease of patients with underlying eye disease. By using a ResNet50, it brings greater efficiency in identifying the fundus images. Representation Learning for Graph Neural Networks Supervisor: LI Xiaomeng / ECE Student: YANG Jingwen / COSC Course: UROP1000, Summer While image shadow detection has obtained significant achievements, its detection over dynamic scenes is yet to be explored until ViSha, the first large scale dataset for video shadow detection. Compared with images, videos have an additional temporal dimension. In this report, we aim to evaluate the effects of the temporal information in video shadow detection. Specifically, we first review the current state-of-the-art video shadow detection method, TVSD-Net, which considers temporal information at video level. Then, we design a modified network that aims to enhance the image-level representations of TVSD-Net by the Swin Transformer. The experiment results show that better image-level representations hardly improve the performance, while the performance of lacking temporal information clearly drops. This inspires us to explore long-term temporal information for continuous research.