The National Institutes of Health (NIH) and the Food and Drug Administration (FDA) are requesting information on what critical resource gaps exist for validation and use of AI/ML to support radiological tool development and clinical data interpretation. This RFI is being released in parallel with a companion RFI (NOT-OD-21-162) focused on resource gaps for Next Generation Sequencing (NGS). If desired, respondents may provide comments that encompass both foci where the fields converge (e.g., linking tumor features with sequencing data, merged datasets). The comment period on this Notice is 90 days. Response to this Notice is voluntary.
Reference materials are needed to facilitate the development, rigorous performance assessment, and validation of AI/ML models (e.g., deep learning) across a full range of clinical radiological applications. A current challenge is the lack of large, ethnically diverse, clinically annotated radiology datasets of sufficient quality with associated metadata that are Findable, Accessible, Interoperable, and Reusable (FAIR) with the appropriate policies and controls in place to ensure responsible data sharing and data use (e.g., privacy protections, consent requirements, compliance with applicable laws and regulations). Data storage and analysis infrastructure and tools for clinical and translational research using radiology data are also needed. Such resource gaps, in general, are frequently identified as limiting factors that impede high-quality research, development, validation, and regulatory science. Addressing these gaps could foster the development and validation of the next generation of AI/ML algorithms (e.g., deep learning and continual learning models) capable of analyzing data from multiple clinical domains (e.g., radiological and NGS) to provide researchers, physicians, and patients with new “big data” insights on the detection, characterization, treatment, and drug resistance of cancers and other diseases.
The NIH and FDA are interested in receiving input on the greatest needs and opportunities for the development of high-quality radiological datasets and tools that can be used to support AI/ML development, particularly in relation to the three topic areas noted below. Since the algorithmic needs for the development and validation of AI/ML algorithms may go beyond the “training” aspect of AI/ML, NIH and FDA would be interested in information related to both training and real-world use of unlocked AI/ML algorithms. NIH and FDA welcome input from research investigators, study participants, professional organizations, and other interested members of the public.
Response Date: November 1, 2021