School of Engineering Department of Electronic and Computer Engineering 136 Robust and Generalized Methods for Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: ZHAO Haihan / CPEG Course: UROP1000, Summer The current project aims to employ the convolutional neural network ResNet-18 to solve the classification task of the CIFAR-10 dataset, a collection of 60000 32*32 RGB images commonly used to train and test typical neural networks, including KNN, linear, and convolutional neural networks. The network was built with PyTorch and executed on Google Colab. This report reviews relevant literature and elaborates on a sample code, including a data loader, a ResNet model, and a train function. The final accuracy is roughly 80%. With the addition of information argumentation and module improvement, it can achieve more satisfactory results and be applied in long-tail recognition tasks. Learning with Limited Anotated Data for Medical Image Diagnosis Supervisor: LI Xiaomeng / ECE Student: GUO Jiarong / ELEC Course: UROP2100, Fall UROP3100, Spring UROP4100, Summer Labeling relatively large datasets always take a significant amount of time, especially in medical images. SemiSupervised Learning (SSL) improves results by utilizing both labeled and unlabeled data. Previous SSL methods assume the labeled and unlabeled data are in a closed set. However, open-world unlabeled data can contain classes of out-of-set data which transfer negative results in unlabeled learning. It is Open-set Semi-Supervised Learning (OSSL). In this paper, we address an exciting but more challenging setting, OSSL and multilabel classification, which is a more realistic setting. We propose a framework for novel class detection and multi-label pseudo-label generation to solve the multi-label semi-supervised learning and open-set problem. We capitalize on the knowledge obtained from labeled data and distribution in feature space. By conducting several experiments on multi-label datasets with retinal and natural images, our methods can significantly improve the model's ability to perform during the evaluation stages. Our code will be released on GitHub.