Scholarlink

Background <p>Accurate preoperative risk stratification of non-muscle-invasive bladder cancer (NMIBC) is essential for developing individualized treatment plans. This study developed an interpretable habitat model based on biparametric MRI to improve preoperative risk stratification in patients with NMIBC.</p> Methods <p>This two-center study enrolled a total of 206 patients. The internal dataset (<i>n</i> = 147) included the training set (<i>n</i> = 81), test set (<i>n</i> = 35), and time-sequential validation set (<i>n</i> = 31). The external dataset included 59 patients. The primary outcome was the dichotomized risk stratification according to the European Association of Urology (EAU) guidelines (low-intermediate risk vs. high-very high risk). K-means clustering was applied to segment tumors into habitat subregions. Following habitat feature extraction and selection, six habitat models were constructed, including support vector machine (SVM), logistic regression, linear discriminant analysis (LDA), Gaussian process classifier, random forest, and multilayer perceptron (MLP). For comparison, a VI-RADS model and a clinical prediction model were also developed. SHapley Additive exPlanations (SHAP) analysis was used to interpret the optimal model and visualize its decision-making process. Furthermore, we conducted propensity score matching (PSM) analysis to evaluate the generalizability of the habitat models.</p> Results <p>All habitat models exhibited good diagnostic performance. The SVM habitat model achieved the highest AUC in both the training set (AUC = 0.95, 95% CI: 0.85–1.00) and the external validation set (AUC = 0.82, 95% CI: 0.61–1.00). It showed a trend toward better performance compared to the clinical model (AUC = 0.68, 95% CI: 0.51–0.83) and the VI-RADS model (AUC = 0.75, 95% CI: 0.58–0.92) in both sets, although the differences did not reach statistical significance. After PSM balancing all baseline covariates (all SMD &lt; 0.1), SVM still maintained stable predictive capacity (AUC = 0.72, sensitivity = 0.769) in matched external cohort, verifying good cross-center generalization.</p> Conclusion <p>Our biparametric MRI-based habitat models, particularly the SVM classifier, showed a trend toward improved preoperative risk stratification of NMIBC. These findings require large multicenter prospective validation before clinical individualized treatment decisions and follow-up 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 validation of a preoperative biparametric MRI-based habitat model for risk stratification of non-muscle-invasive bladder cancer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Pingping He,
  • Ruiquan Chen,
  • Qiaozhen Zheng,
  • Minxuan Tian,
  • Ming Chen,
  • Wei Guo,
  • Ruixiong You

摘要

Background

Accurate preoperative risk stratification of non-muscle-invasive bladder cancer (NMIBC) is essential for developing individualized treatment plans. This study developed an interpretable habitat model based on biparametric MRI to improve preoperative risk stratification in patients with NMIBC.

Methods

This two-center study enrolled a total of 206 patients. The internal dataset (n = 147) included the training set (n = 81), test set (n = 35), and time-sequential validation set (n = 31). The external dataset included 59 patients. The primary outcome was the dichotomized risk stratification according to the European Association of Urology (EAU) guidelines (low-intermediate risk vs. high-very high risk). K-means clustering was applied to segment tumors into habitat subregions. Following habitat feature extraction and selection, six habitat models were constructed, including support vector machine (SVM), logistic regression, linear discriminant analysis (LDA), Gaussian process classifier, random forest, and multilayer perceptron (MLP). For comparison, a VI-RADS model and a clinical prediction model were also developed. SHapley Additive exPlanations (SHAP) analysis was used to interpret the optimal model and visualize its decision-making process. Furthermore, we conducted propensity score matching (PSM) analysis to evaluate the generalizability of the habitat models.

Results

All habitat models exhibited good diagnostic performance. The SVM habitat model achieved the highest AUC in both the training set (AUC = 0.95, 95% CI: 0.85–1.00) and the external validation set (AUC = 0.82, 95% CI: 0.61–1.00). It showed a trend toward better performance compared to the clinical model (AUC = 0.68, 95% CI: 0.51–0.83) and the VI-RADS model (AUC = 0.75, 95% CI: 0.58–0.92) in both sets, although the differences did not reach statistical significance. After PSM balancing all baseline covariates (all SMD < 0.1), SVM still maintained stable predictive capacity (AUC = 0.72, sensitivity = 0.769) in matched external cohort, verifying good cross-center generalization.

Conclusion

Our biparametric MRI-based habitat models, particularly the SVM classifier, showed a trend toward improved preoperative risk stratification of NMIBC. These findings require large multicenter prospective validation before clinical individualized treatment decisions and follow-up planning.