Multimodal Ultrasound Radiomic Technology for Diagnosing Benign and Malignant Thyroid Nodules of Ti-Rads 4-5: A Multicenter Study

Luyao Wang, Chengjie Wang, Xuefei Deng, Yan Li, Wang Zhou, Yilv Huang, Xuan Chu, Tengfei Wang, Hai Li, Yongchao Chen2024 · DOI

TL;DR

This study introduces a multimodal ultrasound radiomic approach to improve the diagnosis of benign versus malignant thyroid nodules, achieving high accuracy across multiple centers. By integrating data from conventional ultrasound, strain elastography, and shear wave elastography, and employing machine learning with explainability features, the model offers a promising tool to aid clinicians and reduce unnecessary biopsies.

What Problem Does This Paper Address?

Diagnosing thyroid nodules accurately is crucial, as misdiagnosis can lead to unnecessary invasive procedures or delayed treatment. While fine-needle aspiration biopsy (FNAB) is the current gold standard, it's invasive and its accuracy can depend on operator skill. Existing ultrasound classification systems like TI-RADS have limitations, and interpreting ultrasound images can be subjective. The paper addresses the need for a more objective, accurate, and non-invasive method to distinguish between benign and malignant thyroid nodules, particularly those classified as TI-RADS 4-5, which have a higher suspicion of malignancy.

Methods

The core of this research lies in a multimodal approach to radiomics. Here's a breakdown:

  • Multimodal Ultrasound Imaging: The study leverages data from four types of ultrasound imaging:
  • Conventional 2D Ultrasound (transverse and longitudinal views)

    • Strain Elastography (SE)
    • Shear Wave Elastography (SWE) This multi-modal approach aims to capture a richer set of information about the nodule's structure, texture, and mechanical properties.
  • Radiomic Feature Extraction: Quantitative imaging features (radiomics) were extracted from these multimodal images. This includes histogram features, and features derived from the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run Length Matrix (GLRLM), and Gray-Level Size Zone Matrix (GLSZM). Wavelet features were also extracted to capture multiscale and multidirectional information.

  • Feature Selection: To manage the high dimensionality of radiomic features and identify the most relevant ones for malignancy prediction, the Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed. This technique helps to reduce overfitting and improve model robustness.
  • Machine Learning Model: A Support Vector Machine (SVM) classifier was used to predict nodule malignancy based on the selected radiomic features. The SVM's hyperparameters were optimized using grid search and 5-fold cross-validation.
  • Model Interpretability: To address the "black-box" nature of machine learning models, the Shapley Additive Explanation (SHAP) method was utilized. SHAP provides insights into which features are most influential in the model's predictions and how they contribute to the final output, allowing for a more transparent understanding of the diagnostic process.

Datasets

The study utilized ultrasound images from 468 patients with thyroid nodules across three medical centers. The data was divided into:

  • Training Cohort (n=312): Used for model training and internal validation.
  • Internal Validation Cohort (n=134): Used for internal testing and model refinement.
  • External Test Cohort (n=62): Used to validate the model's generalization capabilities on data from a different hospital.

The final pathological diagnosis (benign or malignant) based on FNAB or surgical treatment served as the gold standard.

Key Insights and Ideas

  • Multimodality is Key: The study strongly suggests that combining data from different ultrasound modalities (2D US, SE, SWE) significantly enhances diagnostic performance compared to using any single modality alone. This highlights the complementary nature of these imaging techniques.
  • Radiomics Captures Subtle Clues: Radiomic features, particularly those derived from wavelet transforms, proved to be highly effective in distinguishing between benign and malignant nodules. These features capture complex textural and structural information that might be missed by the human eye.
  • Interpretability Builds Trust: The use of SHAP is a critical advancement. It not only explains why the model makes a certain prediction but also helps clinicians understand the underlying features driving the diagnosis, fostering trust and facilitating clinical adoption.
  • Wavelet Features Dominate: The most important features identified by SHAP were all processed through wavelet transforms, underscoring their power in extracting discriminative information from ultrasound images.
  • Clinical Relevance of Features: The features identified by the model align with known clinical indicators for thyroid nodule malignancy, such as nodule edge morphology, internal echogenicity, and hardness distribution, lending credibility to the model's findings.

Key Results

The multimodal radiomic model demonstrated excellent performance:

  • Training Set: AUC of 0.971
  • Internal Testing Set: AUC of 0.856
  • External Validation Set: AUC of 0.779

The combined model consistently outperformed models using single modalities. For instance, in the training set, the combined model achieved an accuracy of 94.6%, sensitivity of 95.4%, and specificity of 92%. The external validation showed an accuracy of 0.806, sensitivity of 0.818, and specificity of 0.714. The SHAP analysis identified five key features, primarily wavelet-based, that were most influential in predicting malignancy.

Limitations and Future Work

The authors acknowledge several limitations:

  • Benign Nodule Imbalance: The dataset had a relatively low number of benign nodules, which could introduce bias. Future work should focus on collecting more benign nodule data to balance the dataset.
  • Retrospective Nature: The study is retrospective, and long-term prospective studies are needed to further validate the model's utility and robustness.
  • Interpretability Nuances: While SHAP provides valuable insights, certain limitations in clinical interpretation standards for AI-assisted diagnosis remain. Future work aims to establish appropriate interpretation standards for all radiologists.

The authors suggest that the model's success in integrating multimodal data and its high scalability make it adaptable for other sensor data processing tasks, potentially extending to multisource data fusion for improved diagnostic accuracy in various medical imaging fields.

Summary updated

BibTeX
@misc{wang2024multimodal,
  author = {Luyao Wang and Chengjie Wang and Xuefei Deng and Yan Li and Wang Zhou and Yilv Huang and Xuan Chu and Tengfei Wang and Hai Li and Yongchao Chen},
  title = {{Multimodal Ultrasound Radiomic Technology for Diagnosing Benign and Malignant Thyroid Nodules of Ti-Rads 4-5: A Multicenter Study}},
  year = {2024},
  doi = {10.3390/s24196203},
}
Figure