Prostate cancer imaging AI benchmark highlights cross-hospital generalization
Published in Artificial Intelligence, Journals, Scanners and Imaging, Tools.
Published October 31, 2026.
A multicentre study published in npj Digital Medicine on October 26 compares AI approaches combining PSMA PET/CT and multiparametric MRI to identify clinically significant prostate cancer.
Chen and colleagues used paired imaging and biopsy-confirmed labels from 527 patients at three hospitals. Models were developed using 363 patients at one centre and evaluated independently in 164 patients from two other centres.
PSMA PET contributed the greatest incremental predictive value. In external testing, PSMA PET/CT achieved an AUC of 0.753, compared with 0.653 for multiparametric MRI. Diffusion-weighted MRI introduced important differences between centres: excluding it improved external AUC by 0.079, alongside variation in acquisition settings. Some simpler pretrained models and late-fusion approaches generalized better externally than alternatives.
This benchmark emphasizes that adding imaging modalities does not automatically improve performance across hospitals. AUC measures discrimination, not the percentage of patients correctly diagnosed. The results do not establish that these models can replace biopsy or improve clinical outcomes. The accepted article remains subject to editorial changes.
Source: Chen et al., npj Digital Medicine, October 26, 2026.