Transparent foundations: post-editing foundation models for explainable and deployable radiology imaging AI

The National Institute of Biomedical Imaging and Bioengineering awarded the University of California, Los Angeles $244,897 on August 18, 2026, under the Discovery and Applied Research for Technological Innovations to Improve Human Health program (CFDA 93.286) to develop and validate AI approaches that enhance transparency and clinical generalizability of foundation models in radiology imaging diagnostics.

The project focuses on lung disease detection on chest X-rays using two specific approaches: integrating clinically interpretable concepts to improve model transparency through concept embeddings, and studying model-editing methods to enhance performance on rare pathologies. The work addresses gaps in understanding medical foundation models by auditing and curating expert-driven concept sets for radiological findings and examining approaches to improve model reliability on long-tailed pathologies, all without requiring full model retraining. The methodologies are designed to be disease and image agnostic, with potential application beyond the initial lung-imaging focus.

Work is performed at UCLA in Los Angeles, California, with a period of performance through July 31, 2028. The award is a Project Grant supporting exploratory and developmental research in response to PA-25-304.

Abstract

Radiology imaging is a cornerstone of modern healthcare, providing accessible and cost-effective diagnostics. The recent integration of deep learning, particularly through foundation models, has increasingly enhanced diagnostic precision, and as of december 2024, over 1,000 AI-enabled medical devices were FDA-approved, with more than 75% are used in radiology. Despite their promise, foundation models face concerns regarding: (1) the lack of transparency in their decision making, which makes IT difficult for clinicians to understand or verify, and (2) limitations in pretraining data with long-tailed pathologies that can affect downstream reliability and generalizability on rare or differential cases. In response to PA-25-304 (exploratory/developmental research), this project will develop and validate new AI approaches to enhance the transparency and clinical generalizability of foundation models in imaging diagnostics without requiring full model retraining. While the focus of this application has been on lung disease detection on chest X-rays, the methodologies can be disease and image agnostic. Based on our preliminary data and our experience with an interdisciplinary team of data scientists and radiologists, we plan to execute specific aims: 1) integrating clinically interpretable concepts to enhance model transparency, and 2) studying model-editing methods to tackle "long-tailed pathologies". The studies proposed in this project are novel and innovative because they address critical gaps in understanding and enhancing the transparency and generalizability of medical foundation models. Aim 1 will be the first to audit and curate expert- driven concept sets for radiological findings and improve model transparency through concept embeddings. Aim 2 will be the first study to examine and enhance performance on rare pathologies, as well as differential diagnoses, through surgical adjustments and few-shot model editing. Our framework offers scalable, actionable, and easily checkable solutions that align foundation models with clinical decision-making processes, ensuring outputs are both transparent and clinically meaningful while maintaining diagnostic utility. The expected outcome of this project is a holistic framework to ensure transparency, explainability, and deployability in foundation models for medical imaging. The success of this project will provide critical insights into radiological risk estimation by (1) enhancing the transparency and explainability of model decision-making, and (2) developing model-editing techniques that maintain diagnostic accuracy across a wide range of pathologies. This project is highly feasible and has the potential to be transformative for both AI-driven data science and clinical medicine.