ProstateMind：多模态前列腺疾病诊断与指南指导管理ProstateMind: Multimodal Prostate Disease Diagnosis and Guideline-Grounded Management
前列腺疾病人工智能系统需要能够适应异质性影像环境，并输出可审计的临床建议。我们开发了 ProstateMind，一个整合自监督前列腺 MRI 表征学习、任务特异性预测模型和检索增强指南推理的多模态临床决策支持系统。基于六个数据来源，包括多中心回顾性队列、公开 MRI 数据集和前瞻性真实世界队列，我们训练了可复用的 3D MRI 编码器，并评估其在前列腺分割、PCa/csPCa 检出、肿瘤负荷、病理升级、精囊侵犯、BPH 功能状态和 Agent 报告生成中的表现。内部测试中，多模态模型在 PCa 和 csPCa 检出中的 AUC 分别为 0.879 和 0.924，外部验证中 csPCa 检出 AUC 为 0.833。深度影像特征签名可实现高阴性预测值的病理升级和精囊侵犯排除。基于指南证据的 Agent 改善了证据检索，生成了专家评价较好的回答，并对 193 例真实世界病例进行了保守分层。ProstateMind 为前列腺疾病多模态管理提供了可审计框架。Artificial intelligence for prostate disease should generalize across heterogeneous imaging settings and return auditable recommendations. We developed ProstateMind, a multimodal clinical decision support system integrating self-supervised prostate MRI representation learning, task-specific prediction models, and retrieval-augmented guideline reasoning. Across six data sources, including multicenter retrospective cohorts, public MRI datasets, and a prospective real-world cohort, we trained a reusable 3D MRI encoder and evaluated prostate segmentation, PCa/csPCa detection, tumor burden, pathological upgrading, seminal vesicle invasion, BPH functional status, and agent-based reporting. Multimodal models achieved AUCs of 0.879 for PCa and 0.924 for csPCa in internal testing, with external csPCa discrimination of 0.833. Deep imaging signatures enabled high-negative-predictive-value rule-out of pathological upgrading and seminal vesicle invasion. The guideline-grounded agent improved evidence retrieval, generated favorably rated answers, and stratified 193 real-world cases conservatively. ProstateMind provides an auditable framework for multimodal prostate disease management.
