MVMAE is a self-supervised radiology framework that combines masked-image reconstruction with cross-view alignment across projections in a study. The repository also includes MVMAE-V2T, a variant that adds paired radiology reports as a text-based training signal while retaining image-only inference.
The code includes commands for MVMAE pretraining, MVMAE-V2T pretraining, and offline evaluation through full fine-tuning or linear probing with an ensemble classifier. The README names MIMIC-CXR, CheXpert, and PadChest as the three public datasets used for downstream disease classification, and includes a PadChest-to-MIMIC label-harmonization CSV.
The repository says the methods outperform supervised and vision-language baselines, with additional V2T gains in low-label settings. It does not provide result tables, model weights, or an independent validation of those claims in the README.