It combines masked image reconstruction with cross-view alignment to learn view-invariant, disease-relevant representations.
The authors also introduce MVMAE-V2T, which uses radiology reports as an auxiliary training signal while retaining vision-only inference. They evaluate downstream disease classification on the public MIMIC-CXR, CheXpert, and PadChest datasets, reporting that MVMAE outperforms supervised and vision-language baselines. MVMAE-V2T reports further gains, particularly in lower-label settings.
The arXiv record lists the first submission on November 27, 2025 and version 5 on October 21, 2026, with a TMLR 2026 journal reference. The paper does not establish clinical safety, prospective performance, regulatory clearance, deployment in care settings, or public availability of a production model.