Upcoming Talks

Towards Robust Medical Image Analysis

Most statistical learning algorithms in medical image analysis rely on an over-simplified assumption, that is, the train and test data are independent and identically distributed. However, in real-world scenarios, it is common for models to encounter data from new and different centers to which they were not exposed during training. This is often the case in medical imaging applications due to differences in acquisition devices, imaging protocols, and patient characteristics. In this work, (i) We collect the first large dataset of T1-weighted and T2-weighted abdominal MRI series from five centers between March 2004 and November 2022 to investigate this robustness challenge. We develop a new pancreas segmentation method, PanSegNet, combining the strengths of nnUNet and a Transformer network with a new linear attention module enabling volumetric computation. (ii) We propose one domain generalization (DG) as a promising direction as it enables models to handle data from previously unseen domains by learning domain-invariant features robust to variations across different domains via adversarial intensity attacking. (iii) To further enhance the segmentation model's robustness, we propose a novel approach to address this challenge by developing controllable diffusion models for medical image synthesis and validating this model's performance on various medical datasets. (iv) Finally, we construct a semi-supervised algorithm to take advantage of the unlabeled medical segmentation data into large-scale medical image segmentation training and show that recent advancements in large models can provide promising solutions to generate robust models against real-world clinical challenges.

Zheyuan Zhang earned his bachelor's degree from Tsinghua University in 2018. During the summer of 2017, he participated in a research internship at Johns Hopkins University. From 2018 to 2025, he pursued a Ph.D. in Biomedical Engineering at Northwestern University, focusing on robust medical image analysis using deep learning methods.

Coming soon

Please add your questions to the speaker either to this google form or directly under the YouTube video