This recommended practice delineates an architecture and offers suggestions for enhancing the generalizability of artificial intelligence (AI) models in medical imaging.

This recommended practice delineates an architecture and offers suggestions for enhancing the generalizability of artificial intelligence (AI) models in medical imaging. It encompasses the following components: A review of AI solutions in medical imaging from a generalizability standpoint; processes and specifications for data processing--including data collection to ensure variety and data annotation to attain pristineness; a means of removing compounding factors to mitigate bias across various datasets; specifications for AI modeling that account for small training data, uni/multi-modal data in medical imaging, thereby enabling flexibility for deployment across distinct usage scenarios; approaches for continuous learning and federated learning to enable AI models to adapt to local populations; and evaluation metrics from a generalizability perspective.

This recommended practice enables stakeholders to develop robust and generalized artificial intelligence solutions in medical imaging in a controlled manner.