Scholarlink

Objective <p>Differentiating between benign and malignant vertebral compression fractures (VCFs) using conventional MRI images presents a significant challenge. This study seeks to develop an MRI-Based Habitat model, to effectively distinguish between benign and malignant VCFs. Additionally, the study aims to employ the SHapley Additive exPlanations (SHAP) method to provide interpretability for the radiomics model.</p> Methods <p>A retrospective analysis of MRI data from 132 patients (ages range 40–97&#xa0;years) with VCFs was performed across two medical centers (January 2021–December 2024), with 97 in the training set and 35 in the external test set. Following the preprocessing and segmentation of the regions of interest in the MRI images, a habitat analysis model was developed utilizing T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI) data. The optimal number of clusters was identified to delineate sub-regions. Subsequently, radiomics features were extracted and filtered, and the Random Forest (RF) algorithm was employed to construct four distinct models: a T1WI classic radiomics model, a T1WI habitat model, a T2WI classic radiomics model, and a T2WI habitat model. The diagnostic performance of each model in distinguishing between benign and malignant VCFs was assessed using the area under the receiver operating characteristic curve (AUC). The model demonstrating the highest performance was further visualized using the SHAP method.</p> Results <p>The analysis determined that the optimal number of clusters for the habitat sub-regions in both T1WI and T2WI sequences was three. Within the external test set, the T2WI habitat model demonstrated superior diagnostic performance, achieving an AUC of 0.919, the T1WI habitat model achieved an AUC of 0.806, and the T1WI and T2WI classic radiomics models AUCs of 0.898 and 0.863, respectively. SHAP value analysis indicated that, within the T2WI habitat model, the absolute weight of Inverse Variance in Habitat sub-region 1 was the highest, with a value of 0.253.</p> Conclusion <p>The T2WI habitat model developed from MRI images demonstrates a robust capability to distinguish between benign and malignant VCFs, surpassing the performance of classic radiomics models. Furthermore, the visualization of SHAP values enhances the model’s interpretability, offering valuable scientific indicators for the early clinical differentiation of benign and malignant VCFs.</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

MRI-based habitat analysis for differentiating benign and malignant vertebral compression fractures

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月2日

查看全文
  • Xiaoshan Liu,
  • Zhiqin Gao,
  • Junqiao Guo,
  • Haifeng Zhang,
  • Hanlin Zhu,
  • Ming Zhan,
  • Bo Feng,
  • Wenyan Wei

摘要

Objective

Differentiating between benign and malignant vertebral compression fractures (VCFs) using conventional MRI images presents a significant challenge. This study seeks to develop an MRI-Based Habitat model, to effectively distinguish between benign and malignant VCFs. Additionally, the study aims to employ the SHapley Additive exPlanations (SHAP) method to provide interpretability for the radiomics model.

Methods

A retrospective analysis of MRI data from 132 patients (ages range 40–97 years) with VCFs was performed across two medical centers (January 2021–December 2024), with 97 in the training set and 35 in the external test set. Following the preprocessing and segmentation of the regions of interest in the MRI images, a habitat analysis model was developed utilizing T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI) data. The optimal number of clusters was identified to delineate sub-regions. Subsequently, radiomics features were extracted and filtered, and the Random Forest (RF) algorithm was employed to construct four distinct models: a T1WI classic radiomics model, a T1WI habitat model, a T2WI classic radiomics model, and a T2WI habitat model. The diagnostic performance of each model in distinguishing between benign and malignant VCFs was assessed using the area under the receiver operating characteristic curve (AUC). The model demonstrating the highest performance was further visualized using the SHAP method.

Results

The analysis determined that the optimal number of clusters for the habitat sub-regions in both T1WI and T2WI sequences was three. Within the external test set, the T2WI habitat model demonstrated superior diagnostic performance, achieving an AUC of 0.919, the T1WI habitat model achieved an AUC of 0.806, and the T1WI and T2WI classic radiomics models AUCs of 0.898 and 0.863, respectively. SHAP value analysis indicated that, within the T2WI habitat model, the absolute weight of Inverse Variance in Habitat sub-region 1 was the highest, with a value of 0.253.

Conclusion

The T2WI habitat model developed from MRI images demonstrates a robust capability to distinguish between benign and malignant VCFs, surpassing the performance of classic radiomics models. Furthermore, the visualization of SHAP values enhances the model’s interpretability, offering valuable scientific indicators for the early clinical differentiation of benign and malignant VCFs.