Multicentre benchmark of PSMA PET/CT and mpMRI fusion for non-invasive clinically significant prostate cancer diagnosis

Clinically significant prostate cancer (csPCa) guides-biopsy and treatment decisions, yet its non-invasive diagnosis remains challenging. The incremental value of PSMA PET/CT beyond multiparametric MRI (mpMRI), and the optimal strategy for multimodal fusion, remain unclear. We established a real-world, multicentre benchmark using 527 patients from three hospitals with fully paired PSMA PET, CT, T2-weighted MRI, and diffusion-weighted imaging (DWI), together with pathology-confirmed biopsy labels. Models were developed in one centre ( n = 363) and independently evaluated in two external centres ( n = 164). We compared frozen pretrained encoders, fine-tuned 2D and 3D CNN and Vision Transformer models, three fusion strategies, and an interpretable cross-modal Transformer. PSMA PET provided the greatest incremental value, with its removal producing the largest AUC reduction internally (−0.077) and externally (−0.055), supported by the highest attention attribution (αPET = 0.45). PSMA PET/CT also outperformed mpMRI externally (AUC 0.753 vs. 0.653). DWI was the principal source of domain shift, as its exclusion improved external AUC by 0.079 amid inter-centre b-value heterogeneity. Intermediate fusion performed most consistently internally, whereas late fusion and pretrained 2D models generalised better externally. This benchmark supports robust, non-invasive csPCa diagnosis by identifying modalities and fusion strategies that generalise reliably across institutions.