Scholarlink

Purpose <p>To evaluate radiomics machine learning (RML) models using different magnetic resonance imaging (MRI) sequences for differentiating posterior fossa A (PFA) and posterior fossa B (PFB) subtypes of posterior cranial fossa ependymoma, classified according to the latest pathological guidelines, and to perform model interpretability analyses.</p> Methods <p>Clinical and radiological data from 124 patients diagnosed with ependymoma in the posterior cranial fossa were retrospectively collected from two separate institutions. Radiomics features were subsequently derived from MRI sequences including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and T2-weighted fluid-attenuated inversion recovery (T2W-FLAIR). Radiomics analyses encompassed three primary phases: feature dimension reduction, construction of feature labels, and statistical differentiation. A prediction model was subsequently developed utilizing logistic regression (LR) with min-max normalization. The performance of this model was evaluated by calculating the area under the curve (AUC), assessing calibration curves (CC), and employing decision curve analysis (DCA). Additionally, Shapley Additive Explanations (SHAP) were utilized to clarify the relationship between radiomics features and biological properties, thereby improving interpretability of the model.</p> Results <p>The diagnostic performance of the fusion model was superior to that of single-sequence models. Specifically, the T1WI + T2W-FLAIR fusion model demonstrated optimal diagnostic efficacy in differentiating PFA and PFB subtypes of posterior cranial fossa ependymoma, with robust internal/external test set (AUC = 0.736/0.756) and high training accuracy (AUC = 0.916). SHAP analysis ranked the nine most discriminative radiomics features from the T1WI + T2W-FLAIR fusion model, reflecting their relative contribution to subtype classification.</p> Conclusion <p>RML models based on multiple MRI sequences can effectively differentiate PFA and PFB subtypes of posterior cranial fossa ependymoma. The T1WI + T2W-FLAIR fusion model provides higher diagnostic accuracy and can offer valuable diagnostic insights for radiologists.</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

A predictive model for differentiating PFA and PFB subtypes of posterior fossa ependymoma using multi-sequence MRI radiomics: a two-center study

  • 期刊论文

  • 发表日期：

    2025年10月23日

查看全文
  • Rui Xu,
  • Hanjiaerbieke Kukun,
  • Jing Xue,
  • Yuhui Xiong,
  • Wei Zhao,
  • Yuwei Xia,
  • Yangyang Li,
  • Pahati Tuxunjiang,
  • Chunhui Jiang,
  • Wenyu Ji,
  • Yunling Wang,
  • Hu Xiao

摘要

Purpose

To evaluate radiomics machine learning (RML) models using different magnetic resonance imaging (MRI) sequences for differentiating posterior fossa A (PFA) and posterior fossa B (PFB) subtypes of posterior cranial fossa ependymoma, classified according to the latest pathological guidelines, and to perform model interpretability analyses.

Methods

Clinical and radiological data from 124 patients diagnosed with ependymoma in the posterior cranial fossa were retrospectively collected from two separate institutions. Radiomics features were subsequently derived from MRI sequences including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and T2-weighted fluid-attenuated inversion recovery (T2W-FLAIR). Radiomics analyses encompassed three primary phases: feature dimension reduction, construction of feature labels, and statistical differentiation. A prediction model was subsequently developed utilizing logistic regression (LR) with min-max normalization. The performance of this model was evaluated by calculating the area under the curve (AUC), assessing calibration curves (CC), and employing decision curve analysis (DCA). Additionally, Shapley Additive Explanations (SHAP) were utilized to clarify the relationship between radiomics features and biological properties, thereby improving interpretability of the model.

Results

The diagnostic performance of the fusion model was superior to that of single-sequence models. Specifically, the T1WI + T2W-FLAIR fusion model demonstrated optimal diagnostic efficacy in differentiating PFA and PFB subtypes of posterior cranial fossa ependymoma, with robust internal/external test set (AUC = 0.736/0.756) and high training accuracy (AUC = 0.916). SHAP analysis ranked the nine most discriminative radiomics features from the T1WI + T2W-FLAIR fusion model, reflecting their relative contribution to subtype classification.

Conclusion

RML models based on multiple MRI sequences can effectively differentiate PFA and PFB subtypes of posterior cranial fossa ependymoma. The T1WI + T2W-FLAIR fusion model provides higher diagnostic accuracy and can offer valuable diagnostic insights for radiologists.