Interpretable multimodal fusion of 18F-PSMA-1007 PET/CT and mpMRI for preoperative ISUP grade prediction in primary prostate cancer

Accurate preoperative International Society of Urological Pathology (ISUP) grading of prostate cancer (PCa) is critical for clinical decision-making. While multi-parametric MRI (mpMRI) is routine, it struggles with occult high-risk lesions due to limited metabolic insights. PET/CT provides complementary molecular information but is associated with higher cost and radiation exposure. This retrospective study included 341 patients (internal cohort) and 36 patients in an independent external validation cohort who underwent preoperative mpMRI, PET/CT, and radical prostatectomy. Postoperative ISUP grade served as the reference. Five binary models were developed: clinical, mpMRI, PET/CT, MPC (mpMRI + PET/CT), and CMPC (clinical+mpMRI + PET/CT). Corresponding three-class models (MPC-3c, CMPC-3c) were also constructed. Performance was primarily evaluated using the area under the receiver operating characteristic curve (AUC). Model interpretability was evaluated using SHAP to quantify clinical feature contributions and Grad-CAM to visualize deep learning decision-making regions. For binary classification in the internal cohort, the clinical, mpMRI, and PET/CT models yielded AUCs of 0.739, 0.881, and 0.888, respectively. Multimodal fusion improved internal performance, with the MPC and CMPC models reaching AUCs of 0.945 and 0.950, respectively. For three-class classification, MPC-3c achieved a macro-AUC of 0.810 (accuracy 0.695), increasing to 0.830 (accuracy 0.698) with CMPC-3c. In the external cohort, biopsy accuracies were 0.743 (binary) and 0.657 (three-class). External binary AUCs for MPC and CMPC were 0.726 and 0.718, respectively, with three-class macro-AUCs of 0.778 and 0.808. These findings provide preliminary external validation but also indicate reduced binary discrimination across centers, warranting cautious interpretation due to the small external cohort. Integrating mpMRI and 18 F-PSMA-1007 PET/CT may improve preoperative ISUP grade prediction by providing complementary anatomical, functional, and molecular information. Clinical features provided complementary information in the internal cohort, whereas external performance remained limited and requires further validation. This decision-support adjunct supplements histopathology and warrants prospective multicenter validation. This retrospectively registered study was registered at the Chinese Clinical Trial Registry on October 14, 2024 (ChiCTR2400090817 http//www.chictr.org.cn/).