Objectives This study evaluated and compared the effectiveness of various predictive models for forecasting occult lymph node metastasis (LNM) in tongue squamous cell carcinoma (TSCC) patients. Methods In this retrospective diagnostic experiment, 268 patients were recruited from three medical centers. Based on the different hospitals from which the patients were recruited, they were divided into a training set, an internal testing set, and two external testing sets, comprising 107, 53, 63, and 45 patients, respectively. Several predictive models were developed using patients’ contrast-enhanced magnetic resonance imaging (CEMRI), including two-dimensional deep learning (2D DL), conventional radiomics (C-radiomics), and intratumoral heterogeneity radiomics (ITH-radiomics). Univariate and multivariate logistic regression analyses were conducted on the clinical data. Finally, two fusion strategies were used to construct the model. Results The ITH-radiomics model exhibited superior discriminative power compared to C-radiomics model. The late fusion model had the highest area under the curve (AUC) across all test sets (0.81–0.85). Compared to the late fusion model, the AUC values for the early fusion, 2D DL, C-radiomics, and ITH-radiomics models in the test sets ranged from 0.77 to 0.82, 0.64 to 0.81, 0.66 to 0.77, and 0.77 to 0.80, respectively. Additionally, the late fusion model demonstrated the highest accuracy (76–89%) and specificity (87–100%) across the test sets. Conclusions The evaluation of the models’ effectiveness revealed that the decision-based late fusion model, which integrated 2D DL, C-radiomics, ITH-radiomics, and clinical data, achieved the best results. This predictive approach can more accurately assess patients’ conditions and aid in selecting surgical plans. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> How well does fusing multiple models work for predicting occult lymph node metastasis in patients with tongue squamous cell carcinoma? <Emphasis Type="BoldItalic">Findings</Emphasis> The late fusion model, incorporating two-dimensional deep learning, conventional-radiomics, intratumoral heterogeneity-radiomics, and clinical features, achieved the best results compared to each individual model. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> Patients with a high intratumoral heterogeneity-radiomics index exhibit an increased risk of occult lymph node metastasis in tongue squamous cell carcinoma patients, which showed that the late fusion model achieves superior predictive performance compared to the early fusion model.

Evaluating fusion models for predicting occult lymph node metastasis in tongue squamous cell carcinoma

期刊论文
发表日期：

2025年3月5日

查看全文

Wen Li,
Yang Li,
Li Wang,
Minghuan Yang,
Masahiro Iikubo,
Nengwen Huang,
Ikuho Kojima,
Yingding Ye,
Rui Zhao,
Bowen Dong,
Jiang Chen,
Yiming Liu

摘要

Objectives

This study evaluated and compared the effectiveness of various predictive models for forecasting occult lymph node metastasis (LNM) in tongue squamous cell carcinoma (TSCC) patients.

Methods

In this retrospective diagnostic experiment, 268 patients were recruited from three medical centers. Based on the different hospitals from which the patients were recruited, they were divided into a training set, an internal testing set, and two external testing sets, comprising 107, 53, 63, and 45 patients, respectively. Several predictive models were developed using patients’ contrast-enhanced magnetic resonance imaging (CEMRI), including two-dimensional deep learning (2D DL), conventional radiomics (C-radiomics), and intratumoral heterogeneity radiomics (ITH-radiomics). Univariate and multivariate logistic regression analyses were conducted on the clinical data. Finally, two fusion strategies were used to construct the model.

Results

The ITH-radiomics model exhibited superior discriminative power compared to C-radiomics model. The late fusion model had the highest area under the curve (AUC) across all test sets (0.81–0.85). Compared to the late fusion model, the AUC values for the early fusion, 2D DL, C-radiomics, and ITH-radiomics models in the test sets ranged from 0.77 to 0.82, 0.64 to 0.81, 0.66 to 0.77, and 0.77 to 0.80, respectively. Additionally, the late fusion model demonstrated the highest accuracy (76–89%) and specificity (87–100%) across the test sets.

Conclusions

The evaluation of the models’ effectiveness revealed that the decision-based late fusion model, which integrated 2D DL, C-radiomics, ITH-radiomics, and clinical data, achieved the best results. This predictive approach can more accurately assess patients’ conditions and aid in selecting surgical plans.

Key Points

Question How well does fusing multiple models work for predicting occult lymph node metastasis in patients with tongue squamous cell carcinoma?

Findings The late fusion model, incorporating two-dimensional deep learning, conventional-radiomics, intratumoral heterogeneity-radiomics, and clinical features, achieved the best results compared to each individual model.

Clinical relevance Patients with a high intratumoral heterogeneity-radiomics index exhibit an increased risk of occult lymph node metastasis in tongue squamous cell carcinoma patients, which showed that the late fusion model achieves superior predictive performance compared to the early fusion model.