Foundation Models for Cancer Imaging Biomarkers
Speaker: Suraj Pai
Affiliation: Mass General Cancer Center
Date: July 2025
Watch the recording on YouTube
Abstract
This seminar will explore our recent methodological breakthroughs in representation learning for developing next-generation foundation models in oncology and radiology. I’ll first discuss a novel foundation model for cancer imaging biomarkers where we trained on over 11,000 diverse lesions through a tumor imaging aware contrastive learning strategy. The model was then evaluated in distinct and clinically relevant applications of cancer imaging-based biomarkers. We found that it facilitated better and more efficient learning of imaging biomarkers and yielded task-specific models that significantly outperformed conventional supervised and other state-of-the-art pretrained implementations on downstream tasks, especially when training dataset sizes were very limited [1]. Next, I will discuss an approach where we broadened this contrastive learning strategy to general radiological understanding in CT scans, trained on 148k publicly available scans. This strategy facilitated learning fine-grained local anatomical representations while outperforming supervised and state-of-the art methods [2]. Finally, I will discuss a handful of studies that leveraged our cancer imaging-based biomarkers to obtain relevant insights and present a benchmark for comparing and evaluating foundation models in their ability to capture radiological cancer phenotypes [3].
- https://www.nature.com/articles/s42256-024-00807-9
- https://arxiv.org/abs/2501.09001
- https://www.researchsquare.com/article/rs-6630446/latest
Speaker bio
Suraj Pai is a researcher at the Mass General Cancer Center, where he is currently building multimodal agentic solutions for radiotherapy treatment planning to be applied into the clinic. He is also stationed at the Artificial Intelligence in Medicine (AIM) Program at Mass General Brigham and Harvard Medical School where his research focuses on developing representation learning methods for cancer imaging and radiology. Previously, he earned an MSc in Artificial Intelligence (cum laude) from Maastricht University and is currently in the last stages of obtaining a PhD. Prior to starting his career as a researcher, he was a Machine Learning Engineer with three years of prior experience in building, maintaining and deploying models in production for consulting and product. Outside of research, he finds inspiration in nature, especially hiking and birding across New England, and enjoys being contemplative through meditation, readings in metaphysics, ethics, epistemology, and world history.
Rights and attribution
This recording and accompanying text are presented for educational access with attribution to the speaker. Speaker views are their own. No open license is implied unless explicitly stated.