The Molecular Imaging Technology Research Program (MITRP) is pleased to announce the publication of a new article in the Journal of Imaging Informatics in Medicine titled “From Embeddings to Accuracy: Comparing Foundation Models for Radiographic Classification”. This work represents a significant partnership between academia and industry, featuring a comprehensive author list from the University of Wisconsin-Madison and Microsoft Health and Life Sciences. The study’s authors from the University of Wisconsin-Madison include Tyler Bradshaw, Richard J. Bruce, John W. Garrett, Xue Li, Dr. Alan B. McMillan, and Joshua D. Warner. Collaborators from Microsoft Health and Life Sciences include Christopher Burt, Noel C. F. Codella, Alexander Ersoy, Matthew P. Lungren, Jameson Merkow, Naiteek Sangani, Alberto Santamaria-Pang, and Ivan Tarapov.
As artificial intelligence continues to reshape medical imaging, the implementation of end-to-end deep learning models remains resource-intensive and often challenges generalizability across different clinical sites. To address these hurdles, the research team systematically evaluated the utility of embeddings derived from pre-trained foundation models—including MedImageInsight, Rad-DINO, and CXR-Foundation—to train lightweight adapter models for multi-class radiography classification. The study focused on the critical clinical task of tube placement assessment and related findings, utilizing a robust dataset of over 8,800 radiographs.
The team found that utilizing MedImageInsight embeddings paired with Support Vector Machine (SVM) or Multi-Layer Perceptron (MLP) adapters yielded a mean area under the curve (mAUC) of 93.1%. This approach notably outperformed a fully fine-tuned Convolutional Neural Network, specifically DenseNet121, while maintaining high computational efficiency with training times measured in just minutes on a standard CPU. Furthermore, fairness analysis on the adapters indicated minimal disparities, with gender differences in performance within 1.8% and consistent results across various age groups. These findings confirm that foundation models can facilitate accurate, efficient, and equitable diagnostic tools, supporting the integration of these methods into clinical workflows where rapid decision-making is essential.
MITRP extends its warmest congratulations to Xue Li and the entire collaborative team for this significant contribution to the advancement of medical imaging AI.
Find the full text of the paper here: https://doi.org/10.1007/s10278-025-01747-5
