UROP Proceedings 2022-23

School of Engineering Department of Computer Science and Engineering 105 Transformers for Medical Imaging and Analysis Supervisor: CHEN, Hao / CSE Student: WANG, Zixuan / COMP Course: UROP1100, Spring Multiple-instance learning has been a classical research topic for a long time, and it is gaining increasing attention, especially in medical image analysis. Whole slide imaging is a relatively new field compared to radiology, and has triggered intensive research. The fact that whole slide images are relatively large, usually gigabytes, has made multiple-instance learning a feasible approach. In this report, I summarized what I have learned through the UROP project, mainly the background knowledge of multiple-instance learning and some state-of-the-art model architecture. I will continue to investigate possible model refinements with implementation and testing in the future. Transformers for Medical Imaging and Analysis Supervisor: CHEN, Hao / CSE Student: YU, Cheuk Hei / COSC Course: UROP1100, Spring Multimodal Learning is an important task in computational pathology, which aims to capture information across modalities. This resembles the scenario in clinics where doctor usually won’t rely on a single source when performing diagnosis. In this work, I was inspired by the idea of domain adaptation and DCGANs, and searched for ways to make the model fail to distinguish between modalities. Specifically, I attempted to formulate multimodal learning as a domain adaptation task and used late fusion techniques to fuse survival prediction scores from each modality. Results, however, are not satisfactory as the tasks have different nature and they are targeted in solving different problems due to some confusions. Artificial Intelligence Against COVID-19 Supervisor: CHEN, Hao / CSE Student: HUANG, Chenliang / COMP Course: UROP1100, Spring This report details my experience with a UROP project exploring the use of machine learning in medical imaging, specifically reproducing MONAI’s COVID-19 Lung CT Lesion Segmentation Challenge baseline. Literature on U-Net and nnU-Net for medical image segmentation tasks and first prize winner Hu’s semisupervised learning model are reviewed, along with potential benefits of using large models. The report also shares lessons learned in implementation, such as the importance of using one package manager and helpful tools like VS Code Remote SSH and Tmux. The project’s implications and value are discussed, emphasizing the importance of cutting-edge technology in medical imaging.