ModaLens measures how medical AI models use images alongside reports
ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Medical AI systems often look at both X-ray images and radiology reports to answer health questions. It is tricky to tell if these systems truly use the images when they already have access to detailed reports. The authors developed ModaLens, a method that swaps images while keeping reports the same, to see how much the AI's answers change. They found the AI relies less on images when reports are available, showing reports reduce the model’s image sensitivity.
What this means in practice
- •For medical ai developers: Evaluate how much radiology AI systems rely on images versus reports using image-swap sensitivity tests.
- •For clinical data teams: Improve trust in AI by auditing models to ensure they appropriately use image data alongside reports.
Authors
Sebastián Andrés Cajas Ordóñez, Maximin Lange, Quang Bui, Anqi Peter Li, Felipe Ocampo Osorio, Rafi Al Attrach, Kushul Reddy Palakala, Sahil Kapadia, Zakaria Laouabdia Sellami, Xinyue Zhang, Ashley Zhang, Leo Anthony Celi
Abstract
A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also uses the image. ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity: MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients, all 14 questions per case (13 finding-specific and one composite), each image replaced by one from another study, usually of the same patient, with question and report fixed. Under an explicit answer instruction, the model's generated answer changes on 4.26 percent of trials with the report and 20.94 percent without it, a paired increase of 16.7 points (patient-clustered 95 percent CI 15.6 to 17.7), so report availability reduces image-swap sensitivity under this protocol; the original prompt with a lowercase first-token readout gives 4.70 percent against 17.07 percent, and substitutions also move continuous answer scores where the binary prediction does not change. The labels are derived from reports, which limits conclusions about visual correctness; the direction replicates in two further model lineages. Code, the exact prompts and a run record for every number are at https://github.com/criticaldata/MODALENS.