EE Seminar: Generative Approaches for Robust Medical Image Segmentation Under Domain Shift and Few-Shot Constraints
הרישום לסמינר יבוצע בתחילת הסמינר באמצעות סריקת הברקוד למודל (יש להיכנס לפני כן למודל, לא באמצעות האפליקציה)
Registration to the seminar is done at the beginning of the seminar by scanning the barcode for the Moodle (Please enter ahead to the Moodle, NOT by application)
Electrical Engineering Systems Seminar
Speaker: Tal Grossman
M.Sc. student under the supervision of Prof. Hayit Greenspan
Sunday, 10th May 2026, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Generative Approaches for Robust Medical Image Segmentation Under Domain Shift and Few-Shot Constraints
Abstract
Medical image segmentation performance often drops under domain shift caused by changes in scanner, protocol, or modality, while collecting new dense annotations is costly. In this thesis, We study whether generative priors can improve robustness in source-free and few-shot settings. We present two complementary methods: SymmAdapt, a source-free adaptation framework based on Symmetrical Flow Matching with round-trip consistency filtering and generative self-training, and DiffuSAM, a prompt-free adaptation method that keeps SAM2 frozen and learns a lightweight diffusion prior over memory embeddings. SymmAdapt achieves strong source-free performance across abdominal, cardiac, and prostate benchmarks, while DiffuSAM reaches competitive few-shot and source-free results with low memory requirements. Together, these results show that generative priors provide domain-invariant semantic structure that supports robust medical image segmentation under realistic clinical constraints.
-סמינר זה ייחשב כסמינר שמיעה לתלמידי תואר שני ושלישי-
This Seminar Is Considered A Hearing Seminar For Msc/Phd Students-
