1. Large language models (LLMs) can produce radiology reports matching radiologist accuracy and outperforming teleradiology preliminaries.
2. Current FDA and EU AI Act criteria are too narrow; broader metrics for reliability and consistency are needed.
3. Proprietary LLMs may transmit patient data externally, creating HIPAA compliance concerns.
4. Commercial LLMs can reflect demographic and racial bias, requiring ongoing bias testing and diverse training data.
5. Local or federated deployment and vendor transparency are essential for responsible clinical integration.

Source: Radiology