School of Engineering Department of Computer Science and Engineering 93 Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LEE Hsin-ning / COMP Course: UROP 3100, Spring Knee osteoarthritis (OA) is a degenerative joint disease primarily affecting elderly individuals, resulting in joint pain and stiffness. Among the causes of knee osteoarthritis, articular cartilage injury plays a significant role. Unfortunately, the regenerative capacity of cartilage is limited, and current treatments are unable to fully restore degraded cartilage in advanced stages. Therefore, early diagnosis of knee osteoarthritis is crucial, and the segmentation of knee articular cartilage on Magnetic Resonance Images (MRIs) can serve as a valuable tool for this purpose. However, the complexity and variability of cartilage structures in MRIs present challenges for effective segmentation and quantification. Moreover, the interpretation of MRIs demands significant time and expertise. In this project, we aim to address these challenges by developing a reliable semi-supervised deep learning model that can accurately segment knee cartilage, enabling objective quantification and grading of cartilage injury. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LI Chengxi / COMP Course: UROP 1000, Summer In the field of Biomedical Image Analysis, traditional CNN and Transformer structures are widely used. However, both structures face the issue of spatial inductive bias, which can lead to over-fitting. This study explores the necessity of spatial inductive bias and proposes a new neural network framework aimed at addressing the over-fitting problem caused by spatial inductive bias in traditional CNN and Transformer neural network. The framework consists of two components: the first part is a purely attention-based neural network which can avoid spatial bias, and the second part is a learnable positional encoding module. The structure effectively avoids over-fitting due to spatial inductive bias while capturing underlying spatial relationships to a certain extent. As the research is still in its preliminary stage, this article mainly introduces the current progress. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LIANG Yan / COMP Course: UROP 1100, Fall UROP 2100, Spring This report presents a comprehensive overview of my UROP2100 research experience. The propose of this project is to explore and leverage different approaches to generate synthetic data, which aims to solve the problem of data imbalance in histopathology. The first section is a brief overview of the pipeline of a general Multiple Instance Learning (MIL) model. We take CLAM as an example to illustrate the pipeline. The second section is about the different generated data methods that I have tried, including the model in the vision area and the model in the NLP area. The third section shows some results of the generated images. The fourth section is about the potential reasons for the unsatisfactory results. Finally, the report concludes with a discussion of the future work and the potential research directions. Due to the limitation of the computing resources, the experiments are not comprehensive, which is also a potential future work.
RkJQdWJsaXNoZXIy NDk5Njg=