Join the Meetup to hear talks from experts on AI, ML, and Computer Vision.

Foundation models like SAM and DINO-v2 have shown strong performance on natural image tasks. However, when applied directly to medical imaging, they often underperform due to domain shifts, limited labeled data, and modality-specific challenges. This raises an important question: how can we adapt foundation models to work reliably and meaningfully in medical images?

In this talk, I will share our research efforts toward answering that question. I will begin by exploring several fine-tuning strategies for different data scenarios, ranging from few-shot labeled examples to large collections of unlabeled scans. These strategies aim to help identify the optimal adaptation framework under various data availability settings. I will then introduce a series of models we developed based on these insights. SegmentAnyBone and SegmentAnyMuscle are two SAM-based models designed for accurate bone and muscle segmentation across all body locations and a wide range of
MRI sequences. MRI-Core is a self-supervised model that learns general-purpose MRI features from unlabeled data and can be easily adapted to multiple downstream tasks.

Finally, I will present a clinical application where one of these models is used to support abdominal surgical risk prediction. This example shows how I have explored using these models to contribute to real-world clinical decision-making. I hope this talk can share some of my experiences in building foundation models that are both practical for research and adaptable to clinical settings and to spark new insights and discussions in this field!

Medical AI models can exhibit concerning biases, such as the ability to predict race from radiology images, which is impossible for human experts. This talk will examine bias and batch effects in medical imaging, beginning with a histopathology case study to illustrate the origins of some of these biases. I'll cover detection methods, such as exploratory data analysis, and mitigation strategies, including careful cross-validation and model-level interventions. While research has shown that foundation models reduce some biases, they don't eliminate the problem entirely. Bias represents a fundamental challenge in medical AI requiring early detection, careful validation, and tailored mitigation approaches.

High-quality data is the cornerstone of effective machine learning in healthcare. This talk presents practical strategies and emerging techniques for managing medical imaging datasets, from synthetic data generation and curation to evaluation and deployment.

We’ll begin by highlighting real-world case studies from leading researchers and practitioners who are reshaping medical imaging workflows through data-centric practices. The session will then transition into a hands-on tutorial using FiftyOne, the open-source platform for visual dataset inspection and model evaluation. Attendees will learn how to load, visualize, curate, and evaluate medical datasets across various imaging modalities.

Whether you're a researcher, clinician, or ML engineer, this talk will equip you with practical tools and insights to improve dataset quality, model reliability, and clinical impact.