Privacy Preserving Performance Analysis of the AI Model Deployed on the MRI Scanner with Multimodal Vision-Language Feedback

Motivation: We propose a privacy-preserving method to monitor any AI-based segmentation model performance without access to ground truth data in real-world applications. Goal(s): Develop a multimodal contrastive vision-language model to analyse segmentation performance while preserving patient privacy. Approach: We trained a multimodal contrastive vision-language model on synthetic text-image pairs to learn correlations, predicting vertebral mask segmentation quality on spine MRI using learned visual and textual embeddings. Results: The model showed excellent correlation between missing vertebrae segmentation images and text descriptions, achieving similarity scores above 0.5 in 97% (571 of 587) and 94% (1278 of 1357) of unique texts in two experiments. Impact: We report a privacy preserving mechanism for monitoring segmentation model performance in terms of simple text logging, rather than quantitative numbers which might require re-interpretation to deduce the performance of the AI model.