Uncertainty-Aware Vision–Language Learning Improves Chest X-Ray Retrieval

Key Points

• The aim is to improve the reliability and performance of vision-language models in biomedical tasks like chest X-ray retrieval.
• Introduced the MedProbCLIP framework for vision-language learning.
• Evaluated on the MIMIC-CXR dataset comparing probabilistic and deterministic methods.
• Assessed performance in retrieval and zero-shot classification tasks.
• MedProbCLIP outperformed deterministic baselines in retrieval tasks.
• Achieved superior performance in zero-shot classification compared to existing methods.
• Significantly improved trustworthiness of embeddings used for retrieval.