UROP Proceedings 2022-23

School of Engineering Department of Electronic and Computer Engineering 134 Artificial Intelligence Methods for Medical Videos Supervisor: LI, Xiaomeng / ECE Student: KU, Ho Ming / ELEC Course: UROP1100, Fall In this report I would mainly talk about things that I have been working on in the past semester, specifically on the topic of medical image classifications with the usage of CLIP. I have tested on various models and prompts on several datasets, and tried to conclude on some insights that might be helpful or meaningful to myself. The report in the followings would first talk about the main objective of the project, the program that I used to test, the methodologies that I use and at last the result that I have got. Artificial Intelligence Methods for Medical Videos Supervisor: LI, Xiaomeng / ECE Student: WU, Qi / COMP ZHANG, Yuyao / DSCT Course: UROP1100, Fall UROP2100, Spring UROP3100, Spring Segment Anything Model (SAM) is a large-scaled foundation model for image segmentation. It has shown great capability in generating masks for multiple natural image segmentation tasks. The model supports various prompts, making it possible to zero-shot transfer to other segmentation tasks like medical image segmentation. Meanwhile, fine-tuning the SAM model to medical domain is also attractive. In this report, we conclude several SAM zero-shot medical image segmentation results and show some feasible way to fine-tune the model on medical datasets. Federated Learning with Medical Images Supervisor: LI, Xiaomeng / ECE Student: FOO, Xiao Yi / CPEG Course: UROP1100, Spring Due to the advancement of technology, specifically development in the medical field, medical imaging has become a vital part to create visual representations of the internal structure of the body. Federated learning is a machine learning method where private devices learn a shared model while keeping all the training data. In traditional machine learning, data is collected and stored in a centralized location which might lead to privacy problems. However, in federated learning method, data is kept privately on devices without being shared, and the global model is trained by the sole communication of parameters. This will greatly help medical imaging as medical institutions cannot combine their data due to privacy regulations.