Zhang Yan, He Shengyun, Chen Jie, Liu Guiqin, Luo Yuansheng, Song Yang, Zhou Yan, Zhu Yinjie, Wei Xiaobin, Wu Guangyu
Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China.
Huangpu Branch, Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
Quant Imaging Med Surg. 2025 Feb 1;15(2):1641-1652. doi: 10.21037/qims-24-1015. Epub 2025 Jan 16.
Although artificial intelligence (AI) algorithms provide reliable prostate volume (PV) measurements across various magnetic resonance imaging devices, their impact on prostate cancer risk stratification for patients with a Gleason score of 6 remains unclear. This study aimed to evaluate the benefits of integrating AI-derived PV and transitional or peripheral zone volume (TZV/PZV) measurements with clinical factors to improve prostate-cancer risk stratification.
Our retrospective cohort included 560 patients with biopsy-confirmed Gleason score 6, stratified based on the outcome of radical prostatectomy as clinically significant prostate cancer (csPC) and clinically insignificant prostate cancer (insPC). We used AI methods for accurate PV and TZV/PZV estimation based on the origin virtual net (Vnet) and with cascade (coarse and fine) Vnet, the best-performing volume segmentation network in the subsequent analysis. We then developed predictive models incorporating clinical factors including age, prostate serum antigen levels, Prostate Imaging Reporting and Data System, positivity in transitional-zone or peripheral-zone biopsy, number of positive cores (foundational model), and novel models integrating PV (model 1) and TZV/PZV (model 2). The efficacy of these models was assessed by the receiver operating characteristic area under the curve (AUC).
For prostate segmentation, the fine cascade Vnet performed best with a Dice similarity coefficient of 0.93 for the whole prostate, 0.82 for the transitional zone, and 0.85 for the peripheral zone. The comparative discriminative power of the three models between csPC and insPC was assessed using the test dataset, indicating an AUC of 0.698 for the foundational model, 0.712 for model 1, and 0.730 for model 2. Model 2 significantly outperformed model 1 (P=0.045) and the foundational model (P=0.005) in distinguishing between csPC and insPC. Model 1 also showed statistically significant improvement over the foundational model (P=0.023).
Incorporating AI-driven PV and TZV/PZV measurements with clinical parameters improves prostate-cancer risk stratification.
尽管人工智能（AI）算法能在各种磁共振成像设备上提供可靠的前列腺体积（PV）测量值，但其对 Gleason 评分为 6 的患者前列腺癌风险分层的影响仍不明确。本研究旨在评估将 AI 得出的 PV 以及移行区或外周区体积（TZV/PZV）测量值与临床因素相结合以改善前列腺癌风险分层的益处。
我们的回顾性队列研究纳入了 560 例经活检证实 Gleason 评分为 6 的患者，根据根治性前列腺切除术的结果将其分为临床显著性前列腺癌（csPC）和临床非显著性前列腺癌（insPC）。我们使用基于原始虚拟网络（Vnet）以及级联（粗粒度和细粒度）Vnet 的 AI 方法来准确估计 PV 和 TZV/PZV，在后续分析中级联 Vnet 是表现最佳的体积分割网络。然后，我们开发了包含年龄、前列腺血清抗原水平、前列腺影像报告和数据系统、移行区或外周区活检阳性、阳性核心数量等临床因素的预测模型（基础模型），以及整合了 PV 的新模型（模型 1）和整合了 TZV/PZV 的新模型（模型 2）。通过曲线下面积（AUC）评估这些模型的效能。
对于前列腺分割，细粒度级联 Vnet 表现最佳，整个前列腺的 Dice 相似系数为 0.93，移行区为 0.82，外周区为 0.85。使用测试数据集评估了三种模型在 csPC 和 insPC 之间的比较判别能力，结果显示基础模型的 AUC 为 0.698，模型 1 为 0.712，模型 2 为 0.730。在区分 csPC 和 insPC 方面，模型 2 显著优于模型 1（P = 0.045）和基础模型（P = 0.0０5）。模型 1 也显示出相对于基础模型有统计学意义的改善（P = 0.023）。
将 AI 驱动的 PV 和 TZV/PZV 测量值与临床参数相结合可改善前列腺癌风险分层。
