The MIMRTL lab is pleased to announce that Xue Li successfully completed her preliminary examination on April 29, 2025. This achievement marks a significant step forward in her doctoral studies, advancing her to Ph.D. candidacy.Xue’s research, under the guidance of her advisor, MIMRTL PI Dr. Alan McMillan, focuses on leveraging representation learning and embeddings to address key challenges in medical imaging analysis. Her preliminary examination committee included Dr. Bill Sethares, Dr. Junjie Hu, and Dr. Pedro Morgado.
Representation learning, a cornerstone of modern machine learning, empowers models to automatically identify and encode meaningful features from complex data, significantly enhancing AI performance across various domains. Embeddings, central to this concept, transform intricate data into structured numerical vectors that capture intrinsic semantic relationships. While embedding methodologies have seen substantial advancements in general AI applications, their full potential remains largely untapped within medical imaging. Current methods often rely on pixel-level loss metrics that may overlook crucial anatomical details, and generative tasks frequently operate in high-dimensional pixel spaces without fully exploiting pretrained embedding models for computational efficiency and clinical relevance. Furthermore, the siloed nature of medical imaging data and associated clinical narratives impedes the development of multimodal embeddings that could provide richer contextual analysis.
Xue’s research project directly tackles these unmet needs by investigating and harnessing embedding-based representations to improve medical imaging approaches. Her work outlined in the preliminary examination is structured around three integrated specific aims. The first specific aim is to develop deep learning frameworks that leverage perceptual (embedding-guided) loss functions to accurately correct attenuation and complete truncation artifacts in breast PET/MR imaging. Her second specific aim focuses on establishing a latent diffusion modeling framework conditioned upon learned embeddings to generate predictive, anatomically consistent optical coherence tomography (OCT) images, thereby facilitating early structural-level assessment of treatment response. Finally, her third specific aim is to leverage pretrained foundation models to derive robust, volume-level embeddings aimed at enabling computationally efficient, embedding-based medical image classification and retrieval tasks.
Collectively, Xue’s interconnected research objectives demonstrate how embedding-based methodologies can meaningfully enhance clinical imaging practices. Her work promises improvements in diagnostic accuracy, computational efficiency, and semantic interpretability, with the potential to translate into transformative clinical advancements in medical imaging workflows.
Please join us in congratulating Xue on reaching this significant milestone in her academic career!