RadZero introduces a similarity-based cross-attention framework that significantly improves vision-language alignment in radiology with zero-shot capabilities.

• Processes high-resolution medical images more effectively than current approaches
• Provides explainable AI decisions through transparent attention mechanisms
• Achieves zero-shot multi-task capability without requiring task-specific training
• Leverages complex radiology reports for more comprehensive learning

This research transforms how AI can interpret radiological images, enabling clinicians to understand model reasoning while supporting multiple diagnostic tasks with a single model—crucial for clinical adoption and trust in medical settings.

RadZero: Similarity-Based Cross-Attention for Explainable Vision-Language Alignment in Radiology with Zero-Shot Multi-Task Capability