Magnetic resonance imaging (MRI) is a non-invasive method of producing detailed images of the body’s internal structure, but acquisition and annotation is labor and time-intensive
Medical Research
October 2022 - June 2023
During my senior year, I began researching in the Magnetic Resonance Research Labs with UCLA Health under Dr. Holden Wu. This work is part of my Bioengineering Capstone Project, where my team’s goal is to develop a deep learning model for knee MRI segmentation, allowing for pathology detection within knee cartilage.
Visual of knee MRI and segmented cartilage label
Our work builds of a Stanford dataset known as SKM-TEA, which is comprised of 155 patients with 160 sagittal knee MRI slices for each patient. The MRIs were are using are known as quantitative double-echo steady state (qDESS), which are quantitative MRIs that acquires 2 sets of images, known as echoes 1 and 2, to construct a T2 map of the MRI images. T2 metrics can give insight to early tissue degeneration. We have built three models: U-Net, Attention U-Net, and R2U-Net. Currently, we are optimizing hyperparameters such as batch size, learning rate, and optimizer while also testing different loss functions such as Dice and Cross-Entropy. Our goal is to build a robust, deep learning model that can segment the different types of cartilage within the knee, which can then be coupled with the T2 maps to identify certain pathological conditions. This pipeline would be faster and more convenient.
©2024 Aaron Li