Francis Deng

I am a radiologist. I support generative AI that improves radiologists’ accuracy, efficiency, and timely care, with meaningful human review of diagnostic interpretations.

AI interpretations should remain preliminary until a radiologist reviews the complete examination, clinical information, and proposed interpretation. This applies to every examination, including AI-designated normal, negative, or high-confidence cases. Meaningful review requires source images, relevant priors, adequate time for independent assessment, and authority to modify or reject output. A cursory review and signature, sampled review, retrospective auditing, nonspecialist acknowledgment, or AI supervision is insufficient.

Mammography computer-aided detection (CAD) offers a cautionary example. FDA’s 2012 M-Vu approval described a retrospective reader study of 21 radiologists and 280 examinations, half cancer-positive, unlike routine screening. Lehman et al. (JAMA Internal Medicine. 2015;175:1828–1837; doi:10.1001/jamainternmed.2015.5231) studied 625,625 digital screening examinations at 66 facilities. CAD did not improve overall diagnostic accuracy; cancer detection was 4.1 per 1,000 with and without CAD. The authors cited annual U.S. CAD spending exceeding $400 million at the time, without established benefit to women: substantial spending without demonstrated clinical value.

Even though genAI promises better performance than traditional CAD, a central principle applies: reader-study performance does not guarantee clinical effectiveness. Because CAD already involved radiologists, the lesson is to preserve meaningful review AND establish that the human–AI workflow works in practice.

Specific answers as relevant to imaging:

Question 1: Risk framework
Include the opportunity to correct errors before clinical reliance, reversibility, and downstream dependence. Distinguish radiologist-assisted interpretation from AI interpretations actionable without specialist review.

Question 2: Directiveness
Assess foreseeable clinical effects, not wording alone. Even preliminary negative results may influence care through reassurance or omission. Clinicians may treat triage outputs as diagnoses despite labeling. Disclaimers or “informational” designations cannot substitute for review.

Question 4: Reviewer expertise
“Healthcare professional review” does not establish competence to interpret images or recognize AI errors. Require a radiologist, or an equivalently trained specialist for the particular examination.

Question 6: Escalation
Assess missed urgent findings, unnecessary alerts, and delays affecting unflagged examinations. Aggregate accuracy can obscure consequential failures. Evaluate whether false-positive alerts and review burden cause fatigue and degrade overall performance. Every examination must remain subject to specialist interpretation.

Question 7: Clinical confirmation
Benchmarking should be followed by confirmation in representative radiologist-assisted workflows. Require evidence of effective error recognition and correction, preserved radiologist control, and meaningful clinical performance. Reader studies alone are insufficient.

Question 12: Measurement
Report examination- and finding-level performance, consequential false negatives and positives, uncertainty, and subgroup results. Separate radiologist–AI from AI-alone performance, and synthetic from real-world results. Safety and effective review matter more than report similarity or speed alone.

Question 13: Synthetic data
Synthetic cases may supplement, but not replace, representative real examinations. In my experience, synthetic images frequently contain unnatural artifacts, raising concerns about generalizability. Require assessment of realism and clinical validity. Synthetic performance cannot replace clinical confirmation or review.

Question 14: Comparators
Compare radiologists with and without AI in representative workflows. Use expert adjudication and clinical follow-up; one radiologist’s report is not infallible gold standard. Assess errors, omissions, recommendations, and downstream consequences. Controlled test performance cannot replace implementation evidence or justify AI-alone interpretation.

Question 15: Alternatives
Use realistic, standard-of-care comparators (eg, specialist review of images). Do not assume no interpretation is the only alternative. Access constraints should not justify unreviewed AI reads. Savings claims should include downstream testing, review burden, and clinical consequences.

Requested FDA position:
Make meaningful radiologist review a condition of intended use of imaging genAI. Foresee off-label uses (eg, CAD-triage being used as CAD-diagnosis). Require clinical implementation evidence (comparison against standard of care, real-world radiologist performance without AI). Preserve review for normal, negative, and high-confidence examinations; neither automation nor retrospective monitoring should replace it.