Scholarlink

Objectives <p>This study developed a composite model that merges radiomic features with clinical indicators, as well as multiple CNN models, to predict recurrent lumbar disc herniation (r-LDH) and facilitate personalized surgical planning and evidence-based perioperative care strategies.</p> Method <p>Among the 573 patients enrolled in the study, the rLDH and negative groups were matched using PSM. A combined model incorporating radiomics and clinical factors was then constructed. Five deep learning predictive models were developed and evaluated using ROC curve analysis, AUC, and other indicators for assessing the performance of classification models.</p> Results <p>After screening and matching, 122 patients were included in the recurrence group and 203 in the control group. Following univariate and multivariate analyses, 4 clinical indicators were selected. Additionally, after radiomic feature extraction and screening, 13 features were ultimately chosen. The rad-score for each patient was calculated based on the coefficients. A combined model was then constructed using the rad-score and clinical indicators, and its AUC in the test set was 0.86. Subsequently, five CNN models were developed. The areas under the curves for the different models in the test set were as follows: MobileNetV3 (AUC = 0.87), ResNet50 (AUC = 0.83), ResNet34 (AUC = 0.79), ResNet18 (AUC = 0.88), and DenseNet121 (AUC = 0.88). Through the DeLong test, along with assessments of performance metrics for classification models, it was demonstrated that among the five models, ResNet18 exhibited superior classification performance compared to the others.</p> Conclusion <p>The radiomics-clinical model and CNN models showed good performance in predicting rLDH after PELD. These approaches can identify imaging features that may not be apparent on routine visual assessment. In clinical practice, the prediction results may help identify patients at increased risk of recurrence, facilitate preoperative counseling, and support closer postoperative follow-up and individualized rehabilitation strategies aimed at improving patient management.</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

Deep learning and radiomics based on MRI predict recurrent lumbar disc herniation following percutaneous endoscopic lumbar discectomy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月30日

查看全文
  • Kangle Song,
  • Deyang Qiao,
  • Xingkun Wang,
  • Xiangzhen Kong,
  • Zhenchuan Liu,
  • Yu Wang,
  • Jilai Wang,
  • Jianlu Wei,
  • Lei Cheng

摘要

Objectives

This study developed a composite model that merges radiomic features with clinical indicators, as well as multiple CNN models, to predict recurrent lumbar disc herniation (r-LDH) and facilitate personalized surgical planning and evidence-based perioperative care strategies.

Method

Among the 573 patients enrolled in the study, the rLDH and negative groups were matched using PSM. A combined model incorporating radiomics and clinical factors was then constructed. Five deep learning predictive models were developed and evaluated using ROC curve analysis, AUC, and other indicators for assessing the performance of classification models.

Results

After screening and matching, 122 patients were included in the recurrence group and 203 in the control group. Following univariate and multivariate analyses, 4 clinical indicators were selected. Additionally, after radiomic feature extraction and screening, 13 features were ultimately chosen. The rad-score for each patient was calculated based on the coefficients. A combined model was then constructed using the rad-score and clinical indicators, and its AUC in the test set was 0.86. Subsequently, five CNN models were developed. The areas under the curves for the different models in the test set were as follows: MobileNetV3 (AUC = 0.87), ResNet50 (AUC = 0.83), ResNet34 (AUC = 0.79), ResNet18 (AUC = 0.88), and DenseNet121 (AUC = 0.88). Through the DeLong test, along with assessments of performance metrics for classification models, it was demonstrated that among the five models, ResNet18 exhibited superior classification performance compared to the others.

Conclusion

The radiomics-clinical model and CNN models showed good performance in predicting rLDH after PELD. These approaches can identify imaging features that may not be apparent on routine visual assessment. In clinical practice, the prediction results may help identify patients at increased risk of recurrence, facilitate preoperative counseling, and support closer postoperative follow-up and individualized rehabilitation strategies aimed at improving patient management.