UM Data Science Research Seminar with Precision Medicine

Schedule

LECTURE 1: 12:00 - 12:30

Speaker: Sina Amirrajab

Subject: Simulation and Synthesis for Medical Image Analysis.

Abstract: A key challenge in developing automated image analysis algorithms, particularly deep learning (DL) models, is the limited availability of high-quality, labeled medical images. This scarcity restricts the ability to train robust and generalizable DL models, which require large and diverse datasets to achieve clinical-level accuracy. 
This research tackles this issue by developing simulation and synthesis frameworks to generate large amounts of artificial medical images with corresponding ground truth labels. Two main approaches are explored: physics-driven image simulation, which relies on the underlying principles of medical image formation (e.g., MRI or CT) to generate artificial data, and data-driven image synthesis, which leverages generative models to learn from real imaging data and produce realistic synthetic images. These methods aim to create rich, diverse datasets that enhance the performance of DL algorithms in tasks such as segmentation, classification, and diagnosis.

LECTURE 2: 12:30 - 13:00

Speaker: Sheng Kuang

Subject: Predicting Homologous Recombination Deficiency and Treatment Responses using a Computed Tomography-based Foundation Model: A Preclinical Study.

Abstract: Homologous recombination deficiency (HRD) is the inability to repair DNA breaks through homologous recombination, leading to genomic instability and increased susceptibility to certain cancers. Deep learning has been used to identify tumor types and mutations, but its application in HRD detection, particularly in imaging data, remains largely unexplored.
This pre-clinical study applied a state-of-the-art foundation model (FM) on computed tomography (CT) images, aiming to: (i) distinguish HRD status in mice with isogenic xenografts, and (ii) predict treatment responses to HRD-targeted therapy. The dataset comprises CT scans of 307 mice with balanced HRD status, collected before and after HRD-targeted/control treatment. The FM demonstrated robust classification performance, achieving an AUC value of 0.88 on the test set, significantly outperforming radiomics and supervised deep learning approaches. Additionally, HRD-related features showed strong predictive ability for short-term and long-term responses to HRD-targeted treatment. Interpretability analysis indicated the important role of texture heterogeneity in HRD classification. 

These findings suggest that the application of FMs can capture HRD-related features in CT and predict treatment response, offering a promising approach to guide therapeutic decision-making.