2026
12:00 PM - 01:00 PM
Tuesday Tue
Location
Medical Physics Seminar - Anders Eklund
Synthetic images for image guided radiation therapy
Time:
12:00pm – 1:00pm Seminar & Discussion
Location:
Zoom Webinar
Webinar Registration:
https://stanford.zoom.us/webinar/register/WN_CGJYoZUpSpi1iZXoXSFPeA
Check your email for the Zoom webinar link after you have registered
Speaker
Dr Anders Eklund, Senior Associate Professor, Linköping University
Anders Eklund received a MSc in applied physics and electrical engineering in 2007, and a PhD in medical informatics in 2012, both from Linköping University, Sweden. During 2012 - 2014 he was a postdoc at Virginia Tech, and he recently completed a sabbatical visit at Stanford. He is currently a senior associate professor at Linköping University, focusing on developing methods for medical image processing, statistics, federated learning, radiotherapy, neuroimaging and synthetic image generation.
Abstract
Automatic segmentation of tumor and organs at risk can accelerate radiation therapy treatment planning from hours to minutes. However, training deep learning-based segmentation models normally requires diverse and large annotated datasets. Regulations like HIPAA and GDPR prevent sharing ofsensitive medical data to create large datasets. In this presentation I will therefore explain how generative AI can be used to create realistic synthetic medical images which can be used to train segmentation models. Specifically, I will briefly explain generative adversarial networks (GANs) and diffusion models, and show that segmentation models can indeed be trained with synthetic images. As synthetic images do not belong to a specific person, they can potentially be shared freely as GDPR does not apply to anonymized data. I will also present the problem of memorization in generative AI, i.e. that a generative model can learn to simply generate an image from the training set. The presentation will end by discussing different ways to prevent memorization.