Prediction of Glioma Molecular Markers Using MRI-based Deep Learning
To Know
About this Class
This presentation reviews the evolving role of artificial intelligence in medical imaging, with emphasis on neuro-oncology applications that extend beyond current clinical triage tools toward noninvasive virtual biopsy of gliomas. After introducing machine learning and deep learning concepts, the talk focuses on radio-genomic prediction of clinically important molecular markers including IDH mutation, 1p/19q codeletion, and MGMT promoter methylation. Central to this effort is the curated UTSW-Glioma dataset, including a TCIA-hosted collection of 625 adult preoperative glioma cases with multi-contrast MRI, molecular marker data, metadata, and high-quality tumor segmentations, developed to support robust model training and validation. The presentation further describes deep learning approaches for IDH and 1p/19q classification, confidence score estimation, and clinical deployment through an end-to-end workflow, while addressing persistent challenges of data scarcity, imbalance, and multi-institutional data sharing. Finally, it presents generative AI strategies, including latent diffusion models, alternatives to federated learning, and feature-guided latent diffusion for anatomically consistent 3D MRI editing, enabling tumor insertion, molecular status manipulation, tumor removal, and data augmentation for synthetic image generation without task-specific retraining.