Nuclear Medicine AI
Use this nuclear medicine AI guide to evaluate PET and SPECT reconstruction, image enhancement, segmentation, quantification, interpretation support, and post-deployment performance.
Nuclear medicine AI can support PET and SPECT reconstruction, denoising, segmentation, quantification, interpretation, and outcome prediction. The useful question is not whether a model improves an image. It is whether that improvement helps the clinical task without weakening quantitative consistency, reader confidence, or patient safety.
For buyers and nuclear medicine teams, the safest review starts with the tracer, scanner, protocol, reconstruction method, clinical task, and reference standard. A result from one combination should not be assumed to transfer to another.
About Nuclear Medicine AI
Where AI Fits in Nuclear Medicine
| Workflow stage | AI role | What to verify |
|---|---|---|
| Acquisition and reconstruction | Denoising, deblurring, attenuation correction, motion correction, or shorter acquisition support | Scanner, tracer, dose, acquisition time, reconstruction settings, and task-level image quality |
| Segmentation | Organ, lesion, or region-of-interest boundaries | Boundary accuracy, correction burden, and stability across sites and disease burden |
| Quantification | Uptake values, volumes, kinetic measures, or change over time | Repeatability, calibration, reference method, and version consistency |
| Interpretation support | Detection, classification, comparison, or prioritization | Clinical endpoint, false positives, false negatives, reader interaction, and intended use |
| Therapy planning and follow-up | Dosimetry support, response measurement, or longitudinal comparison | Required human review, uncertainty, patient population, and monitoring after updates |
PET and SPECT Need Task-Based Validation
The SNMMI RELAINCE guidelines recommend evaluating nuclear medicine AI by the clinical task and stage of evidence. Technical image similarity alone is not enough. A cleaner-looking image must still preserve the signal needed for detection, classification, quantification, or another intended decision.
The guidelines describe evidence that moves from early promise to technical task-specific efficacy, clinical decision-making value, and post-deployment performance. That progression gives a buyer a practical test: ask what claim the available evidence can actually support.
Image Enhancement Can Change Quantification
AI-based PET and SPECT enhancement may support lower dose, shorter scans, better signal-to-noise ratio, or improved spatial resolution. Those gains can be valuable, but enhancement can also alter features that a reader or quantitative workflow depends on. Review should include lesion detectability, uptake measurements, reconstruction consistency, and performance across scanner and protocol differences.
When an AI-generated image looks sharper, compare the clinically important result, not only the appearance. For longitudinal care, confirm that the new processing does not create a false change between studies.
Questions for Local Acceptance Testing
- Define the exact task. Name the tracer, modality, scanner family, protocol, anatomy, patient group, and clinical decision.
- Test local data. Include common cases, difficult cases, low-count studies, artifacts, and the range of scanners and protocols in use.
- Measure correction work. Track how often users edit segmentation, reject an output, repeat processing, or spend extra time reviewing a result.
- Protect quantification. Check repeatability and agreement for the measurements that influence care.
- Plan monitoring. Record the product version and watch performance after scanner, protocol, tracer, or software changes.
Integration and Workflow
Ask where processing occurs, how studies are routed, when output returns, how results appear in PACS or a specialty workstation, and what happens during downtime. Confirm whether derived images and measurements enter the medical record, whether users can see the original data, and how corrections are preserved.
The responsible nuclear medicine physician remains accountable for the final interpretation. AI output should be presented as reviewable support with a clear intended use and a defined response when the output conflicts with clinical judgment.
Procurement Questions
- Which tracers, scanners, reconstruction settings, and patient groups are within the intended use?
- What technical, clinical, external, and prospective validation is available?
- How does performance change with low-count data, artifacts, small lesions, or uncommon disease?
- What measurement agreement and repeatability results are available?
- How are model updates versioned, disclosed, and revalidated?
- What local infrastructure, interfaces, storage, and support are required?
- Which metrics will the site monitor after launch?
Sources and Next Steps
- Journal of Nuclear Medicine: RELAINCE Evaluation Guidelines
- Journal of Nuclear Medicine: AI for PET and SPECT Image Enhancement
- ACR Imaging AI Practice Parameter Announcement
Continue with Clinical Validation, Implementation, and the Medical Imaging AI ROI Calculator.