Emory University
Introduction of various datasets available for Datathon
This session will demonstrate how to leverage pre-trained large-scale models for specific classification tasks, covering techniques such as fine-tuning, prompt engineering, and transfer learning to optimize performance on medical diagnosis, image recognition, and other categorical prediction problems while addressing domain-specific challenges and data requirements.
Introduction to foundation models, understanding tokens and their usage, with demonstrations of zero-shot learning and visual question answering (VQA).
Practical session on embedding extraction and multimodal applications using MedGemma, integrating data from CXR, mammography, and more.
This session will explore how to adapt large-scale pre-trained models for pixel-level and region-based segmentation tasks, covering techniques for medical image segmentation, object detection, and semantic parsing while addressing challenges in model adaptation, annotation efficiency, and performance optimization for precise boundary detection and classification.
Overview includes LLM-generated labels, automatic segmentation, CXR preprocessing, and quality control.
17-24 August ’25
Atlanta, Georgia, United States
©2025, HITI Lab. All rights reserved.
