School of Engineering Department of Electronic and Computer Engineering 142 Learning with Limited Anotated Data for Medical Image Diagnosis Supervisor: LI Xiaomeng / ECE Student: LIANG Jialin / ELEC Course: UROP 1000, Summer Medical image segmentation is a popular part of the field of Artificial Intelligence. And Unet is put forward as a State-of-the-Art model that significantly increases training performance. So, the target of this project is to conduct the whole training process with Unet on my own. Through this project, I obtained a brief understanding of the process of medical image segmentation, which consists of three major components: dataset pre-processing, training, and testing. Each facet plays a vital role in shaping the final model. With hands-on experience, I complete various tasks to improve my understanding of the entire procedure. This report aims to provide a comprehensive overview of the tasks completed, analyzing knowledge I learned from this project. Visual-Language Large Foundation Models and Their Applications in Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: LI Jiaxiu / CPEG Course: UROP 1100, Summer Liver is one of the most common and deadly sites for cancer and metastasis. Automated liver imaging analysis provides a solution to the potential increase in the number of liver cancer patients, and the lack of consistency and reproducibility for human labelling. After the recent global shortage of contrast agent, existing solutions are further expected the robustness to CT image contrast level. This report introduces the development of methodologies in the field of automated liver segmentation, including deep neural networks and datasets for biomedical application. It further briefs the ongiong TriALS challenge, and evaluates the application of a proposed nnU-Net-based model in the challenge. Ultra-low Latency Network Transport for Real-Time Video Streaming Supervisor: MENG Zili / ECE Student: PARK Hyun Hu / COMP Course: UROP 1100, Spring Real-time applications are tremendously emerging in recent years. From video conferencing to virtual reality and quantitative trading, real-time applications require an extremely low latency of < 100 ms or even lower. Operators for decades have deployed content delivery networks, edge access nodes and numerous architectures to reduce the RTT to be as low as 10-20 ms. Also, some applications (e.g., WebRTC) choose to rely on UDP sockets since UDP offers more flexibility in loss recovery and congestion control. However, compared to the TCP sockets, using UDP sockets brings considerable development and maintenance overhead to the operators. For UDP sockets, the operators need to maintain everything (e.g., congestion control, loss recovery) themselves. But when using TCP sockets, one function send() will handle all transport functionalities since they are all in the Linux kernel. In the meantime, when new congestion control algorithms are proposed, operators have to re-implement the algorithm again to fit the algorithm into their own frameworks. Consequently, content providers either need to maintain a separate team for the different protocol stack in WebRTC, or start to go back to TCP sockets.
RkJQdWJsaXNoZXIy NDk5Njg=