Multiphasic CT-Based Radiomics Analysis for the Differentiation of Benign and Malignant Parotid Tumors

Qiang Yu, Anran Wang, Jinming Gu, Quanjiang Li, Youquan Ning, Juan Peng, Fajin Lv, Xiaodi ZhangFrontiers in Oncology 2022 · DOI

TL;DR

This study leverages multiphasic CT scans and machine learning to differentiate benign from malignant parotid tumors, achieving high accuracy. The findings suggest that radiomics, particularly when combined with clinical data and analyzed across multiple CT phases, offers a powerful, non-invasive tool for improving diagnostic decisions in parotid gland tumors.

What Problem Does This Paper Address?

Parotid gland tumors are common, but accurately distinguishing between benign (BPTs) and malignant (MPTs) types before surgery is challenging. Current methods like fine-needle aspiration carry risks and can be inaccurate, especially for deep-seated tumors. While CT and MRI are used, their reliance on subjective "semantic" features often misses subtle tumor heterogeneity. This gap leads to suboptimal treatment planning and patient outcomes, as benign and malignant tumors require vastly different surgical approaches and follow-up care. This paper aims to bridge this gap by exploring the potential of radiomics, a quantitative imaging analysis technique, to provide a more objective and accurate differentiation.

Methods

The researchers employed a retrospective study design, collecting multiphasic enhanced CT images from 312 patients with pathologically confirmed parotid tumors. They divided the data into training (n=218) and testing (n=94) sets.

Key methodological aspects include:

  • Radiomics Feature Extraction: Features were extracted from plain, arterial, and venous phases of the CT scans using the PyRadiomics software. This included first-order statistics, shape-based features, texture features, and wavelet features.
  • Data Balancing: To address class imbalance (more benign than malignant tumors), the Synthetic Minority Oversampling Technique (SMOTE) was applied to the training set.
  • Feature Selection: Three distinct methods were used to select the most discriminative radiomics features:
  • Least Absolute Shrinkage and Selection Operator (LASSO)
    • Mutual Information (MI)
    • Recursive Feature Extraction (RFE)
  • Machine Learning Models: Two classifiers were paired with the selected features:
  • Support Vector Machine (SVM)

    • Logistic Regression (LR) This resulted in 24 radiomics models (3 feature selection methods x 2 classifiers x 4 phases: plain, arterial, venous, and three-phase combination).
  • Clinical and Combined Models: For comparison, a clinical model using clinical-radiological features (size, scope, shape, margin, location, infiltration of surrounding tissue, and lymphatic metastasis) was built. A combined model integrating selected radiomics features and clinical features was also developed.

  • Model Evaluation: Performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, and accuracy. Calibration curves were used to evaluate model reliability.

Datasets

The study utilized a dataset of 312 patients with parotid tumors from the First Affiliated Hospital of Chongqing Medical University, collected between January 2014 and October 2021. This dataset comprised 205 cases of benign parotid tumors (BPTs) and 107 cases of malignant parotid tumors (MPTs). The data was split into a training set (n=218) and a test set (n=94). No new benchmark datasets were introduced; the study focused on analyzing existing CT data.

Key Insights and Ideas

  • Radiomics Outperforms Clinical Features: The study strongly suggests that radiomics features, which capture quantitative tumor heterogeneity invisible to the naked eye, provide significantly better discrimination between benign and malignant parotid tumors compared to traditional clinical-radiological features alone.
  • Multiphasic CT is Crucial: Analyzing CT images across multiple phases (plain, arterial, and venous) yields superior diagnostic performance than using any single phase. This highlights the importance of dynamic contrast enhancement in revealing tumor characteristics.
  • LASSO-SVM is a Winning Combination: The LASSO feature selection method combined with the SVM classifier consistently produced the best-performing radiomics models, especially when applied to multiphasic CT data.
  • Morphology Matters: Sphericity, a shape-based radiomics feature, was found to be significantly different between BPTs and MPTs, with benign tumors exhibiting more regular morphology. This underscores how subtle geometric differences can be diagnostically relevant.
  • Wavelet Features are Powerful: Wavelet features, which capture spatial heterogeneity across various scales, played a significant role in the radiomics models, indicating their utility in characterizing complex tumor structures.
  • Combined Models Offer Complementary Value: While the best radiomics model (LASSO-SVM, three-phase) was highly effective, combined models integrating clinical and radiomics features also showed excellent performance, suggesting that a synergistic approach can further enhance diagnostic accuracy, particularly by incorporating extratumoral information like lymphatic metastasis.

Key Results

The study established 28 prediction models in total.

  • Optimal Radiomics Model: The LASSO-SVM model based on a three-phase combination achieved the highest performance in the test set, with an AUC of 0.936 (95% CI = 0.866-0.976), sensitivity of 0.78, specificity of 0.90, and accuracy of 0.86.
  • Comparison to Clinical Model: This optimal radiomics model demonstrated significantly better prediction performance than the best clinical model (LR-based), which had an AUC of 0.781 (p = 0.012).
  • Single-Phase Radiomics: Even single-phase radiomics models, particularly those from the arterial phase, showed strong performance (e.g., LASSO-SVM arterial phase: AUC = 0.924, sensitivity = 0.78, specificity = 0.89).
  • Combined Model Performance: The LR-based combined model achieved an AUC of 0.933 in the test set, comparable to the best radiomics model.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-Center Study: The study was conducted at a single institution, which may limit the generalizability of the findings.
  • Small Malignant Tumor Cohort: The number of malignant parotid tumors was relatively small, which could impact the robustness of the models for this specific class.
  • Internal Validation Only: The models were validated using an internal test set. External validation with prospective, multicenter studies is needed to confirm their generalization ability.

Future work suggested includes:

  • Deep Learning Integration: Exploring deep learning models, which can process image information more efficiently, for multiclassification tasks of parotid tumors.
  • Multiomics Combinations: Investigating the integration of radiogenomics, radiopathomics, and multiradiomics approaches for a more comprehensive understanding and diagnosis of parotid tumors.
  • External Validation: Conducting multicenter prospective studies to validate the developed models.

Summary updated

Abstract

Multiphasic CT-based radiomics analysis showed a machine learning model based on clinical-radiological features and radiomics features has the potential to provide a valuable tool for discriminating benign from malignant parotid tumors.
BibTeX
@inproceedings{yu2022multiphasic,
  author = {Qiang Yu and Anran Wang and Jinming Gu and Quanjiang Li and Youquan Ning and Juan Peng and Fajin Lv and Xiaodi Zhang},
  title = {{Multiphasic CT-Based Radiomics Analysis for the Differentiation of Benign and Malignant Parotid Tumors}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.913898},
}
Figure