School of Engineering Department of Computer Science and Engineering 102 Data-Efficient, Domain Generalizable and Interpretable Deep Learning Supervisor: CHEN Hao / CSE Student: WANG Ziming / COMP Course: UROP 1100, Spring Biomedical image analysis based on deep learning has gained increasing importance in the interdisciplinary field of healthcare and academic research. However, facing challenges across the inconsistency in data labels, modalities, and benchmarks, it isn’t easy to generalize the performance of machine learning algorithms on different analysis tasks. Focusing more on the machine learning part amidst all other contributions in endto-end systems, the MedMNIST v2 dataset provides a simple yet efficient benchmark that preserves the primary data modalities while at the same time standardizing all formats. The purpose of my work this semester is to evaluate the diagnostic performance of ResNet18 and ResNet50 trained on PathMNIST, ChestMNIST, DermaMNIST, and RetinaMNIST with sizes of 28x28 and 224x224. Additionally, Gradientweighted Class Activation Mapping is used as a tool to visualize the learning results, providing a transparent and explainable decision process. Through these experiments, it is shown that MedMNIST v2 is a lightweight dataset that enables models to learn the characteristics of different diseases efficiently. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: CHAN Chun Hin / COMP Course: UROP 1100, Spring Hip osteoarthritis (HOA) is one of the most prevalent musculoskeletal diseases happen in middle-aged or elder adults. With no current available cure treatment, it is better to diagnosis this disease as early as possible. Recent studies highlight the potential of using Deep Learning (DL) for prediction, though evidence is still emerging. In our study, we first use SEResNet50 and EfficientNet B8 to perform classical classification on the presence of HOA by given a 3D MRI image. We then leveraged multiple-instance learning (MIL) to diagnose the presence of HOA based on a bag of 3D MRI images from a patient with ResNet18. We found out that SEResNet50 performs better than EfficientNet B8, and ResNet18 can achieve a recall of 0.70 in test set.
RkJQdWJsaXNoZXIy NDk5Njg=