AI & Machine Learning in Nuclear Medicine¹⁸F-FDG

Quick Answer

Of all the imaging specialties, nuclear medicine may be the one AI is poised to change most — not just to read faster, but to remake the workflow. The reason is structural: nuclear medicine data is quantitative and numeric (SUV, counts, kinetics — signal a model can learn on, not just pictures), it is volumetric and multimodal (PET + CT/MR), it is paired with hard ground truth (pathology, response, survival), and — uniquely — it feeds a closed see→treat→quantify→reassess loop that generates exactly the labeled, outcome-linked data machine learning is hungry for. Deep learning is already contributing across that pipeline: denoising and deep-learning reconstruction (scans at lower dose or shorter time without losing quality), automated detection and segmentation (whole-body tumor volume in seconds, not an hour of manual contouring), automated dosimetry (the step that makes personalized radioligand-therapy dosing practical), and radiomic prognosis. The single biggest unlock is dosimetry: automating it converts fixed-activity, one-size-fits-all therapy into personalized, dose-optimized treatment at scale. The honest counterweight — kept front and center below — is that AI inherits real problems of generalization, bias, and explainability, so the realistic trajectory is profound augmentation with the physician still central, not replacement.

The revolution: why this field, specifically

Every imaging specialty is being touched by AI, but nuclear medicine has four features that make the impact deeper here than in anatomic radiology. First, its output is already a number — models train on quantitative signal, so the leap from "detect" to "measure" to "predict outcome" is short. Second, the theranostic loop is a self-labeling data engine: image the target, treat it, measure the delivered dose, watch the response — every patient generates paired imaging-and-outcome data, the scarce resource that limits medical AI elsewhere. Third, the field's rate-limiting manual step — voxel dosimetry — is precisely what AI automates well; removing that bottleneck is what turns personalized RPT from a research luxury into routine care, arguably the most consequential change coming to the specialty. Fourth, total-body/long-axial-FOV PET plus deep-learning denoising collapse dose and scan time so far that molecular imaging edges toward screening-scale accessibility. Layered on top of all this is a workforce shortage in an exploding therapy field (see the Therapy Transformation) — which makes an AI force multiplier not a threat to jobs but a necessity to meet demand.

Where it already helps

The nearest-term, most concrete applications:

Why NM is fertile ground

Three properties make molecular imaging a strong ML substrate. First, it is inherently quantitative — SUV, kinetic parameters, and counts are numbers, so models learn on signal, not just appearance. Second, studies are volumetric and multimodal (PET + CT/MR), giving rich input. Third, nuclear medicine sits close to ground-truth endpoints — biopsy, therapy response, survival — so models can be trained and validated against outcomes that matter. Total-body/long-axial-FOV PET compounds this by producing dynamic, whole-body, low-dose datasets ideal for data-hungry methods.

The honest limits

A tool is only as good as its validation in your setting. Key cautions: generalization — a model trained on one scanner/reconstruction/population may degrade on another (domain shift); bias — skewed training data produce skewed performance, potentially worsening care for under-represented groups; explainability — a black-box output is hard to trust or override at the point of care; and automation bias — the human tendency to defer to a confident algorithm. Responsible use means understanding a tool's training population, intended use, and failure modes, keeping a human in the loop, and treating regulatory clearance as a floor, not a guarantee of performance in your patients.

What it means for the field

For residents, fluency with these tools is becoming part of the craft — knowing when a denoised or AI-segmented output can be trusted, and when to revert to source data. For elite practitioners and program builders, AI is the enabler that makes personalized dosimetry and high-throughput quantitative reads practical — directly supporting the therapy transformation. For students weighing the field, it is part of why nuclear medicine is a frontier: a specialty where molecular signal, quantitative computing, and targeted therapy converge. The trajectory is augmentation — faster, lower-dose, more reproducible, more personalized — with the physician's judgment still central.

High-Yield Pearls

• Nuclear medicine may be the imaging field AI transforms most, because its data is quantitative-native and its theranostic loop is a self-labeling data engine.
• The biggest single unlock is automated dosimetry — turning fixed-activity therapy into personalized, dose-optimized treatment at scale.
• NM suits ML because it is quantitative, volumetric/multimodal, and outcome-linked.
• Nearest-term wins: denoising/DL-reconstruction (lower dose, faster), auto-segmentation (tumor volume), and automated dosimetry.
• MRAC and synthetic attenuation maps let PET/MRI quantify without CT.
• Real limits: generalization/domain shift, bias, explainability, automation bias — validate in your setting, keep a human in the loop.
• The realistic trajectory is augmentation, not replacement of interpretation.

Common Pitfalls

• Deploying a model outside its validated population/scanner and trusting the output.
• Automation bias — deferring to a confident algorithm over discrepant source data.
• Treating regulatory clearance as proof of performance in your specific patients.
• Ignoring training-data bias and its effect on under-represented groups.

Related Pages

• Context: Total-body & long-axial-FOV PET, Emerging & research PET tracers, SUV harmonization & EARL; vision: Why nuclear medicine — the future of imaging.

Self-Check

Key References

• Reviews of deep-learning reconstruction, denoising, segmentation, and dosimetry in PET/SPECT.
• Literature on AI generalization, bias, and explainability in medical imaging; regulatory frameworks for clinical AI tools.