“We want to build a physicsinformed neural network, so we need to decide how to include this prior physics knowledge,” she explains. “We can’t understand how this deep neural network works, so we can’t manually make this deep neural network work the same as the physics equation. That’s the main challenge.” The breakthrough came when Juyeon employed the encoderdecoder framework, allowing the decoder to use MRI physics to regularize the target model. Integrating physics prior knowledge into the deep neural network improves generalization over previous methods and helps bridge the gap between theory and practice in medical imaging. “It’s hard to get a lot of medical training data for MRF, with ground truth, quantitative values,” she points out. “Usually, people train a model with a synthetic dataset and test it on a real dataset. That’s what we did also, and we improved generalization performance. We believe that by using our techniques, hopefully, it can have better results on real MRI data for patients to improve personalized treatment and medical diagnosis.” Looking ahead, Juyeon intends to extend her research by considering spatial information, which has the potential to improve it further. Additionally, she recognizes that her work on MRF is just one piece of the puzzle. She hopes to explore other aspects of the process and whether the whole process works for real-world cases. “MRF is so interesting because it can be really helpful for medical diagnosis,” she says. “I think this is so valuable. I’mfascinated by how we can include prior knowledge in deep neural networks, so it’s so interesting to put our welldeveloped physics information into the deep neural network.” Originally from Seoul in South Korea, Juyeon embarked on her academic journey at Cambridge two years ago when she started her PhD. “I like it so much,” she smiles. “There are a lot of other PhD students living together. Also, I live in the college, so it’s nice to interact with them and focus on the study.” 4 DAILY MICCAI Wednesday Poster Presentation
RkJQdWJsaXNoZXIy NTc3NzU=