UROP Proceedings 2022-23

School of Engineering Department of Computer Science and Engineering 95 AI + Healthcare: Research and Development of Intelligent Systems for Medical Diagnosis and Applications Supervisor: CHAN, Gary Shueng Han / CSE Student: ZHU, Boan / CPEG Course: UROP2100, Spring This semester, I mainly focus on the paper named “Deep Learning Enables Automatic Detection and Segmentation of Brain Metastases on Multisequence MRI”. In this paper, it proposed a problem which states that the segmentation performance for small tumors is worse than those bigger tumors. This problem can be very serious since this may have a significant influence on early detection of diseases. Therefore, I did some experiments to verify this result and tried to find some ways to solve this problem. In this report, I will first summarize the paper “Deep Learning Enables Automatic Detection and Segmentation of Brain Metastases on Multisequence MRI”, explain the experiments I did, and put forward some ideas. Neural Rendering Supervisor: CHEN, Qifeng / CSE Student: NG, Pui Him Aidan / COMP Course: UROP1100, Spring The performance of generative Artificial Intelligence (AI) models improved significantly in recent years. Commercial applications using these tools to accomplish different tasks, such as conversion and text-toimage generation, gained enormous popularity. In this project, we studied several generative AI models and looked into their basic mechanisms, strengths, and weaknesses. We also explored an important concept related to language models, prompt engineering. With a fundamental understanding, we proposed ways to utilize these generative tools better. In particular, we combined a chatbot with a text-to-image model to enhance image generation. Hoping to leverage AI to help students, we also developed a prototype that leverages a language model to assist them in schedule planning. Neural Rendering Supervisor: CHEN, Qifeng / CSE Student: TSE, Wai Chung / COSC Course: UROP1100, Spring Probabilistic Diffusion Models have become increasingly popular in recent years due to their ability to generate highly realistic images and even video clips. It applies two processes: a forward diffusion process and a reversed process to corrupt and recover data image. The ability of these probabilistic diffusion models is an interesting topic to investigate. In this report, we will conduct an investigation on the generative ability of diffusion models. We first introduced the evolution of generative models, then we performed prompt testing for two of the latest generative models: Stable Diffusion and Midjourney. Then we discussed some important characteristics of the generated image with future improvements.