AI in nuclear medicine

aka deep learning PET/SPECT, aPROMISE

Computer software trained on thousands of scans can now automatically outline tumours on PET and SPECT images in seconds instead of the 30-45 minutes it takes a doctor, though a specialist still checks the result.

Artificial intelligence, predominantly convolutional neural networks and related deep-learning architectures, is being applied across nuclear medicine and radioligand therapy for automated lesion detection and segmentation, tumour-burden quantification, image denoising and reconstruction, CT-free attenuation/scatter correction, and early-stage dose-prediction and toxicity models. In PSMA PET, deep-learning segmentation tools (for example the aPROMISE platform) have matched inter-radiologist agreement for prostate-gland segmentation (Dice similarity coefficient around 0.89 versus an inter-radiologist DSC of about 0.91) while completing a scan in roughly 15 seconds compared with 30-45 minutes for manual contouring, and a comparative reader study against a standard PACS viewer found similarly high intra- and inter-reader agreement with both systems against histopathology as the reference standard, with the AI tool automatically segmenting about 92% of consensus lesions.

A 2026 structured narrative review across PSMA, somatostatin-receptor/PRRT, radioiodine and 90Y-radioembolisation platforms found the strongest direct evidence for AI-assisted segmentation, tumour-burden quantification and early dose-map prediction, but concluded that no platform yet has a prospectively validated, outcome-linked model for AI-guided treatment selection or activity prescription; independent, multicentre, external validation against dosimetric reference standards remains the key evidence gap. Test-retest repeatability studies of AI-derived PSMA tumour-volume metrics show acceptable reproducibility when the same tracer is used for both scans, but substantially wider limits when different PSMA tracers (for example 68Ga-PSMA-11 versus 18F-PSMA-1007) are compared, underscoring that AI outputs, like human-read metrics, are tracer- and protocol-dependent. Current consensus across systematic reviews is that AI should be used as an adjunct that improves efficiency and reproducibility, with expert nuclear medicine physician oversight remaining necessary given model 'black box' limitations and performance variability.

Generic schematic · not to scale · placeholder for the front

How it works

Neural networks trained on large sets of scans with expert-drawn reference contours learn to recognise the imaging patterns of disease directly from pixel data, so a new scan can be automatically segmented, quantified or flagged in seconds by pattern-matching against what the network learned, rather than a person tracing every lesion by hand.

Strengths

Limitations

Since

Deep-learning segmentation tools for PET/SPECT emerging from the mid-2010s; clinical deployment of platforms such as aPROMISE from the early 2020s

Dependencies

Open the dependency map →

Depends on (2)

Quantitative imaging biomarkers (RECIST, PERCIST, SUV, ADC)PSMA PET

Needed by (0)

Nothing in the corpus depends on this yet.

Dependencies are what this technology cannot be delivered without: manufacturing steps, instruments, software, upstream methods. See its full chain on the map.