SPRINTR-PSMA-PET 

PI Elin Trägårdh

A multi-tracer AI model for PSMA PET/CT

Automated tumour detection and burden quantification in prostate cancer

PSMA PET/CT is central to the staging and restaging of prostate cancer, but reading whole-body scans and outlining every lesion by hand is slow and varies between observers. We are developing an AI model that automatically detects and segments prostate-cancer lesions, primary tumour or recurrence, lymph-node metastases, and bone metastases, and quantifies the total tumour burden.

What makes it different is that the model is built to work across tracers and across hospitals. It is trained on three PSMA tracers ([18F]PSMA-1007, [68Ga]PSMA-11 and [18F]DCFPyL) using data from Skåne University Hospital, London Health Sciences Centre (Canada) and the public Auto-PET dataset, and then externally validated on independent cohorts from Umeå, within the SPRINTR study, as well as Turku (Finland) and Heidelberg (Germany).

At each centre, two nuclear-medicine physicians segment the lesions independently. The AI is benchmarked against them on lesion-level sensitivity and positive predictive value, and on tumour-volume agreement (Bland–Altman and Spearman analysis), with inter-reader variability as the reference. The aim is a robust, generalisable tool for reproducible tumour-burden quantification - a step toward AI-supported risk prediction in high-risk prostate cancer and biochemical recurrence.