A Transformer-Based Model Integrating Intratumoral Habitats and Peritumoral Radiomics for Detecting Pelvic Lymph Node Metastasis in Prostate Cancer.
Authors
Affiliations (10)
Affiliations (10)
- Department of Radiology, Nanchong Central Hospital/The Second Clinical Medical College of North Sichuan Medical College, Nanchong, China (J.C., H.L., J.Y.).
- Department of Interventional Medicine Center, The Second People's Hospital of Yibin, Yibin 644000, Sichuan, China (X.F.).
- Department of Radiology, The Third Hospital of Mianyang, Sichuan Mental Health Center, Mianyang 621000, China (R.L., F.L.).
- Department of Radiology, North China University of Science and Technology Affiliated Hospital, Tangshan 063000, China (T.L.).
- Department of Radiology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan 430022, China (L.Y.).
- Department of Radiology, Luzhou Traditional Chinese Medicine Hospital, Luzhou 646000, China (F.W.).
- Department of Radiology, The fifth hospital of Deyang, Deyang 618099, China (P.L.).
- Department of Radiology, Leshan Hospital, Chengdu University of Traditional Chinese Medicine, Leshan 614000, China (Y.X.).
- Department of Medical Imaging Center, The Second People's Hospital of Yibin, 96# Beida Street, Yibin 644000, Sichuan, PR China (Y.F.).
- Department of Radiology, The Third Hospital of Mianyang, Sichuan Mental Health Center, Mianyang 621000, China (R.L., F.L.). Electronic address: [email protected].
Abstract
Pelvic lymph node metastasis (PLNM) is a critical factor in prostate cancer (PCa) treatment decisions. Current imaging and clinical nomograms remain limited by suboptimal sensitivity and frequent underdiagnosis. This study aimed to develop and validate a transformer-based model integrating intratumoral habitat and peritumoral radiomics features for noninvasive preoperative PLNM prediction. A retrospective cohort of 867 PCa patients from four centers who underwent radical prostatectomy and pelvic lymph node dissection was enrolled. Patients were split into training (n = 437), internal validation (n = 125), and external test (n = 305) cohorts. Radiomic features were extracted from tumor habitats and peritumoral rings (3/6/9 mm). Unimodal models were constructed and fused using a transformer architecture that combined habitat, optimal peritumoral, and clinical variables. Performance was assessed using AUC, calibration curves, and decision curve analysis (DCA). Feature importance was interpreted via SHAP values. The habitat model outperformed all unimodal models (AUC 0.788-0.834) and both radiologists (5+ and 10+ years' experience), followed by the 6-mm peritumoral model (AUC: 0.729-0.835). The fusion model achieved superior performance across cohorts (AUC: 0.824-0.917; accuracy: 0.797-0.840; sensitivity: 0.869-0.939) and demonstrated good calibration (P > 0.05). DCA confirmed greater net clinical benefit. Performance remained robust across T-stage and Gleason Grade Group subgroups. The transformer-based fusion model offers accurate, sensitive, and interpretable prediction of PLNM, reducing underdiagnosis and overdiagnosis and supporting individualized clinical decision-making.