Scholarlink

Background <p>Preoperative discrimination between early (N0/N1) and advanced (N2/N3) nodal stages is critical for treatment planning in tongue squamous cell carcinoma (TSCC) due to its well-established impact on prognosis. Nevertheless, conventional imaging techniques show limited accuracy in staging precision. To address this gap, we developed and validated a Deep Learning Radiomics (DLR) model utilizing multimodal MRI, with the goal of enhancing preoperative staging reliability and supporting clinical decision-making in oral oncology.</p> Methods <p>A total of 579 patients with TSCC from two independent centers were enrolled. Patients from Center 1 were randomly split into training (<i>n</i> = 357) and internal test (<i>n</i> = 153) sets, while patients from Center 2 formed an external test set (<i>n</i> = 69). Handcrafted radiomic features were extracted from primary tumor volumes on contrast-enhanced T1-weighted and T2-weighted images. Concurrently, deep learning features were extracted from the maximum cross-sectional region of interest using three convolutional neural networks (ResNet34, ResNet101, DenseNet121). These features were integrated to build the DLR models. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), with comparisons made using DeLong’s test. Calibration and decision curve analysis were also employed.</p> Results <p>ResNet34 DLR model achieved optimal performance, with AUC values of 0.811 (95% CI: 0.76–0.860), 0.806 (95% CI: 0.709–0.903), and 0.807 (95% CI: 0.684–0.931) in the training, internal validation and external test sets, respectively. It significantly outperformed the conventional radiomics model in both the training and internal validation cohorts (all <i>p</i> &lt; 0.05). Models based on ResNet101 and DenseNet121 also consistently surpassed the radiomics baseline across all datasets.</p> Conclusion <p>The proposed DLR approach outperformed conventional radiomics in preoperative prediction of nodal stages in TSCC, demonstrating superior diagnostic potential based on multimodal MRI. This method may offer a non-invasive decision-support tool for individualized treatment planning in oral oncology, although further prospective validation is warranted.</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 radiomics based on multimodal MRI for preoperative prediction of N stage in tongue squamous cell carcinoma: a multicenter study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月30日

查看全文
  • Xin Wang,
  • Sen Fu,
  • Jie Ren,
  • Pingqing Tan,
  • Tao Wei,
  • Yikang Liu,
  • Hang Ling

摘要

Background

Preoperative discrimination between early (N0/N1) and advanced (N2/N3) nodal stages is critical for treatment planning in tongue squamous cell carcinoma (TSCC) due to its well-established impact on prognosis. Nevertheless, conventional imaging techniques show limited accuracy in staging precision. To address this gap, we developed and validated a Deep Learning Radiomics (DLR) model utilizing multimodal MRI, with the goal of enhancing preoperative staging reliability and supporting clinical decision-making in oral oncology.

Methods

A total of 579 patients with TSCC from two independent centers were enrolled. Patients from Center 1 were randomly split into training (n = 357) and internal test (n = 153) sets, while patients from Center 2 formed an external test set (n = 69). Handcrafted radiomic features were extracted from primary tumor volumes on contrast-enhanced T1-weighted and T2-weighted images. Concurrently, deep learning features were extracted from the maximum cross-sectional region of interest using three convolutional neural networks (ResNet34, ResNet101, DenseNet121). These features were integrated to build the DLR models. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), with comparisons made using DeLong’s test. Calibration and decision curve analysis were also employed.

Results

ResNet34 DLR model achieved optimal performance, with AUC values of 0.811 (95% CI: 0.76–0.860), 0.806 (95% CI: 0.709–0.903), and 0.807 (95% CI: 0.684–0.931) in the training, internal validation and external test sets, respectively. It significantly outperformed the conventional radiomics model in both the training and internal validation cohorts (all p < 0.05). Models based on ResNet101 and DenseNet121 also consistently surpassed the radiomics baseline across all datasets.

Conclusion

The proposed DLR approach outperformed conventional radiomics in preoperative prediction of nodal stages in TSCC, demonstrating superior diagnostic potential based on multimodal MRI. This method may offer a non-invasive decision-support tool for individualized treatment planning in oral oncology, although further prospective validation is warranted.