School of Engineering Department of Electronic and Computer Engineering 162 Robust and Generalized Methods for Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: LI Chunsheng / ELEC Course: UROP 1000, Summer Robust and generalized methods includes a lot and the main method is federated learning. Federated learning is emerging as a promising machine learning technique in the medical field for analyzing medical images, as it is considered an effective method to safeguard sensitive patient data and comply with privacy regulations. Aimed at spreading the related details about federated learning with medical images, this project attracts new students to learn some basic knowledge about deep learning and the application in the medical area. Thus, when the new students joined in this project, they got a chance to learning the federated learning by doing an important assignment. And there are many interesting and challenging things in this assignment. Visual-Language Large Foundation Models and Their Applications in Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: LI Jiaxiu / CPEG Course: UROP 2100, Spring The application of retrieval-augmented generation (RAG) models in biomedical and clinical natural language processing (NLP) offers a promising way to improve model performances in high-risk medical settings. This paper presents a survey of medical-specific RAG systems, focusing on their core architecture, domain adaptation methods and integration with structured biomedical knowledge. It begins with an overview of RAG models, then examines key adaptations needed for medical applications and reviews leading RAG models developed for this purpose. The paper also proposes a framework to guide future research on building reliable, domain-specific generative models for medicine. Visual-Language Large Foundation Models and Their Applications in Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: LIU Yilong / CPEG Course: UROP 1000, Summer This report summarizes my learning journey and project progress in the UROP 1000 program, where I explored deep learning techniques for medical image analysis. Starting with foundational knowledge from a CNN course taught by Andrew Ng on Coursera and Python programming basics from Xuefeng Liao’s tutorials, I implemented progressively complex projects: (1) a simple MNIST digit classifier achieves 96.5% test accuracy in 10 epoch. (2) a U-Net model for image segmentation which incorporated skip connections, LeakyReLU activations, and dropout layers handles biomedical data complexities. And (3) a Skin Lesion Classification task using ResNet50 with weighted sampling to address class imbalance. Additionally, I studied a CLIP-based model for retinal disease diagnosis, though its implementation faced challenges. The report details methodologies, experiments, and key insights from each project, highlighting both successes and unresolved issues. Future work will focus on refining the CLIP model, resolving the abnormal training loss in skin lesion classifier task and expanding 3D-Unet applications to the Spleen CT Segmentation problem.
RkJQdWJsaXNoZXIy NDk5Njg=