Scholarlink

Purpose <p>To develop and validate a preoperative [18F]PSMA-1007 PET-derived deep learning score (DLS) and an integrated model combining DLS, D’Amico risk classification, and SUVmax for predicting biochemical recurrence-free survival (BRFS) after radical prostatectomy (RP).</p> Materials and methods <p>This retrospective study included 697 patients with prostate cancer who underwent preoperative [18F]PSMA-1007 PET/CT before RP at three campuses within one healthcare network: Qingchun (training cohort, <i>n</i> = 445), Yuhang (validation cohort 1, <i>n</i> = 190), and Zhijiang (validation cohort 2, <i>n</i> = 62). Five convolutional neural networks were trained to predict 3-year biochemical recurrence from PET images, and the output of the best-performing network was defined as the DLS. Four preoperative prognostic models were evaluated: DLS alone, D’Amico classification alone, D’Amico plus SUVmax, and D’Amico plus SUVmax plus DLS. Performance was assessed using Harrell’s C-index, inverse-probability-of-censoring-weighted time-dependent AUC, calibration, decision curve analysis, and nested model comparisons.</p> Results <p>VGG19 showed the best fixed-time classification performance, with 3-year ROC AUCs of 0.834, 0.755, and 0.723 in the training and two validation cohorts, respectively. A higher DLS was associated with shorter BRFS in all cohorts. Biopsy ISUP Grade Group was significantly associated with BRFS (global <i>P</i> &lt; 0.001). In multivariable analysis, intermediate- and high-risk D’Amico groups, SUVmax, and DLS were independently associated with BRFS. The integrated model achieved C-indices of 0.846, 0.806, and 0.774 and 36-month AUCs of 0.853, 0.801, and 0.779 in the training and two validation cohorts, respectively, outperforming D’Amico classification alone and D’Amico plus SUVmax.</p> Conclusion <p>A fully preoperative model integrating an [18F]PSMA-1007 PET-derived DLS with D’Amico risk classification and SUVmax improved prediction of BRFS after RP. The model may support preoperative risk counseling and postoperative surveillance planning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development and multicohort external validation of a preoperative [18F]PSMA-1007 PET-derived deep learning score and multimodal model for predicting biochemical recurrence-free survival after radical prostatectomy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月11日

查看全文
  • Tiancheng Li,
  • Nina Xu,
  • Xiaofang Yan,
  • Guolin Wang,
  • Zhenfeng Liu,
  • Yinuo Liu,
  • Kui Zhao,
  • Xinhui Su

摘要

Purpose

To develop and validate a preoperative [18F]PSMA-1007 PET-derived deep learning score (DLS) and an integrated model combining DLS, D’Amico risk classification, and SUVmax for predicting biochemical recurrence-free survival (BRFS) after radical prostatectomy (RP).

Materials and methods

This retrospective study included 697 patients with prostate cancer who underwent preoperative [18F]PSMA-1007 PET/CT before RP at three campuses within one healthcare network: Qingchun (training cohort, n = 445), Yuhang (validation cohort 1, n = 190), and Zhijiang (validation cohort 2, n = 62). Five convolutional neural networks were trained to predict 3-year biochemical recurrence from PET images, and the output of the best-performing network was defined as the DLS. Four preoperative prognostic models were evaluated: DLS alone, D’Amico classification alone, D’Amico plus SUVmax, and D’Amico plus SUVmax plus DLS. Performance was assessed using Harrell’s C-index, inverse-probability-of-censoring-weighted time-dependent AUC, calibration, decision curve analysis, and nested model comparisons.

Results

VGG19 showed the best fixed-time classification performance, with 3-year ROC AUCs of 0.834, 0.755, and 0.723 in the training and two validation cohorts, respectively. A higher DLS was associated with shorter BRFS in all cohorts. Biopsy ISUP Grade Group was significantly associated with BRFS (global P < 0.001). In multivariable analysis, intermediate- and high-risk D’Amico groups, SUVmax, and DLS were independently associated with BRFS. The integrated model achieved C-indices of 0.846, 0.806, and 0.774 and 36-month AUCs of 0.853, 0.801, and 0.779 in the training and two validation cohorts, respectively, outperforming D’Amico classification alone and D’Amico plus SUVmax.

Conclusion

A fully preoperative model integrating an [18F]PSMA-1007 PET-derived DLS with D’Amico risk classification and SUVmax improved prediction of BRFS after RP. The model may support preoperative risk counseling and postoperative surveillance planning.