School of Engineering Department of Electronic and Computer Engineering 141 Federated Learning with Medical Images Supervisor: LI Xiaomeng / ECE Student: YUAN Mingxuan / ELEC Course: UROP 1000, Summer During this summer, I participated in the UROP project and conducted research on medical image segmentation under the leadership of Professor Li Xiaomeng and Dr. Wang Haonan. In one month, I learned the basic principles and models of machine learning, such as CNN, VNET, etc. And learned how to simply apply principles to build models, and how to learn and try to reproduce excellent works on GitHub. Throughout the learning process of the entire project, I learned a lot about computer programming. Through reading code and hands-on practice, I effectively strengthened my proficiency and practical ability in programming. Federated Learning with Medical Images Supervisor: LI Xiaomeng / ECE Student: YUAN Ye / ELEC Course: UROP 1100, Fall I focused on three parts during this semester. The first part is replicating FedLC baseline and testing the performance of FedLC with a few encoder-decoder stage uploaded to server. The second part is the ELEC4010N assignment, which contains two parts, write classification task for EMNIST and MNIST datasets and write classification and segmentation task from scratch for ISIC and MRI datasets. The third part is trying to implement classification method on segmentation to exceed baseline. I’ve tried to implement pFedVem and FedVAC. Learning with Limited Anotated Data for Medical Image Diagnosis Supervisor: LI Xiaomeng / ECE Student: LI Hongrui / COMP Course: UROP 1000, Summer Convolutional Neural Networks have a wide range of applications in both natural and medical image processing, and specifically, the Unet and its enhanced versions Unet++, nnUnet and Attention-Unet have been proved quite powerful in 3D medical image processing. In this UROP1000 project, we first implemented different networks on datasets of 3D CT and MRI scans to complete classification and segmentation tasks and got acceptable results. However, in 3D medical images, since volume-wise labeling is quite a demanding and time-consuming task, there is always a short of labeled data, which limits the application of fullysupervised learning in 3D medical image segmentation. Hence, the second stage of our project focused on exploring the utilization of semi-supervised learning in volumetric medical image segmentation.
RkJQdWJsaXNoZXIy NDk5Njg=