Foundation Models in Imaging: From Pneumonia Risk Stratification to Tomographic Reconstruction

by Dr Amir Ehsan Khorashadizadeh

Thursday 21 May 2026, 11:30 → 13:00 Europe/Zurich

OHSA/B17

Abstract: 
Recent advances in foundation models have shown how large-scale pretraining can significantly improve medical image analysis, especially in low-data settings. In this talk, I will first present a pneumonia risk stratification study based on 3D chest CT imaging in collaboration with Kantonsspital Aarau. We show that features extracted from a pretrained segmentation foundation model combined with a simple linear classifier significantly outperform strong convolutional neural networks trained from scratch. This demonstrates the strength of transferable representations learned from large and diverse datasets. Motivated by these findings, the talk then explores a broader question relevant to computational and scientific imaging: can foundation models also transform tomographic image reconstruction? I will present our ongoing work on developing a foundation model for tomographic reconstruction that learns generalizable and scalable priors across datasets and imaging geometries, enabling efficient adaptation to downstream reconstruction tasks.

TEAMS link