CT-based radiomics with various classifiers for histological differentiation of parotid gland tumors

Yang Lu, Haifeng Liu, Qi Liu, Siqi Wang, Zuhui Zhu, Jianguo Qiu, Wei XingFrontiers in Oncology 2023 · DOI

This study, published in Frontiers in Oncology, explores whether machine learning applied to non-contrast CT scans can accurately distinguish between different types of parotid gland tumors, potentially reducing the need for invasive biopsies or expensive contrast-enhanced imaging.

TL;DR

Researchers developed a radiomics pipeline using non-contrast CT images to differentiate four types of parotid gland tumors. While the models performed exceptionally well at identifying malignant tumors and distinguishing most benign types (AUCs up to 0.90), they struggled to differentiate Pleomorphic Adenoma (PA) from Basal Cell Adenoma (BCA), revealing that the "best" machine learning classifier varies depending on the specific diagnostic task.

What Problem Does This Paper Address?

Parotid gland tumors—the most common salivary gland tumors—are a diagnostic headache. While 80% are benign, the surgical approach varies wildly depending on the histological type. For example, a Warthin tumor (WT) might only require conservative treatment, whereas a Pleomorphic Adenoma (PA) requires partial removal due to its higher recurrence and malignancy risk.

Currently, doctors rely on fine-needle aspiration (which is operator-dependent) or contrast-enhanced MRI/CT (which carries risks of allergic reactions to contrast agents and higher costs). This paper investigates radiomics: the process of extracting thousands of quantitative features from standard medical images that are invisible to the human eye. Specifically, the authors wanted to see if non-contrast CT—a faster, cheaper, and safer modality—could provide enough "hidden" data to classify these tumors accurately.

Methods

The researchers employed a classic radiomics-to-machine-learning pipeline:

  1. Segmentation: Radiologists manually outlined the tumor boundaries (Region of Interest) on 3D-Slicer software.
  2. Feature Extraction: Using the PyRadiomics package, they extracted 1,323 features per patient, ranging from simple shape and intensity to complex textures and wavelet-transformed data.
  3. Feature Selection: To avoid "overfitting" (where a model memorizes noise instead of patterns), they used a three-step filter:
  4. ICC: Keeping only stable features.
  5. t-test: Keeping features with statistical significance between groups.
  6. LASSO: A regression method that shrinks less important feature coefficients to zero, leaving only the most predictive "signatures."
  7. Classification: They tested five mainstream classifiers: Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), K-Nearest Neighbors (KNN), and General Bayesian Network (Gnb).

Datasets

The study used a retrospective dataset from a single center (Third Affiliated Hospital of Soochow University):

  • Total Patients: 249.
  • Tumor Types: Pleomorphic Adenoma (71), Warthin tumor (68), Basal Cell Adenoma (41), and Malignant Parotid Gland Tumors (69).
  • Hardware: Data came from four different CT scanners (GE, Siemens, Toshiba, and Philips) to ensure the model wasn't biased toward one manufacturer's image quality.
  • Split: 70% for training and 30% for independent testing.

Key Insights and Ideas

  • No "Silver Bullet" Classifier: One of the most important takeaways is that no single algorithm won every category. For instance, Random Forest was best at identifying malignancies (MPGTs vs. PA), but Logistic Regression was superior for distinguishing between the two most common benign types (PA vs. WT). This suggests that clinical AI tools might need to be "ensembles" of different models rather than a single fixed algorithm.
  • The Power of Wavelets: The vast majority of the "winning" features were wavelet-transformed. These features decompose the image into different frequencies, capturing subtle texture patterns that represent the underlying biological heterogeneity of the tumor.
  • The PA vs. BCA Hurdle: The model hit a wall when trying to distinguish PA from BCA (AUC 0.602). The authors hypothesize this is due to their similar pathological origins; both contain complex mixtures of glandular and myoepithelial cells, making their "radiomic signatures" nearly identical on non-contrast CT.

Key Results

The models showed high diagnostic accuracy for most pairwise comparisons:

  • PA vs. WT: Achieved an AUC of 0.902 and 88% accuracy using Logistic Regression.
  • WT vs. MPGTs: Achieved an AUC of 0.893 using SVM.
  • WT vs. BCA: Achieved an AUC of 0.861 and a staggering 94% accuracy using Random Forest.
  • Malignancy Detection: The models generally performed well (AUCs > 0.83) at separating malignant tumors (MPGTs) from benign ones.
  • The Failure Case: PA vs. BCA remained difficult, with an AUC of only 0.602 and 68% accuracy.

Limitations and Future Work

The authors acknowledge several limitations:

  1. Single-Center Bias: The data comes from one hospital. Multi-center studies are needed to prove the model works globally.
  2. Sample Size: While 249 is a decent cohort for this niche, the BCA group was relatively small (41 cases), which may have limited the model's ability to learn its specific features.
  3. Missing "Phases": The study only used non-contrast CT. Future work could incorporate arterial and venous phase contrast CT to see if that provides the missing information needed to solve the PA vs. BCA puzzle.
  4. Recurrence Prediction: A promising next step is using these radiomic signatures to predict which tumors are likely to recur after surgery.

Summary updated

Abstract

Our study demonstrated that noncontrast CT-based radiomics could stratify refined pathological types of parotid tumours well but could not sufficiently differentiate PA from BCA. Different classifiers had the best diagnostic performance for different parotid tumours. Our study findings add to the current knowledge on the differential diagnosis of parotid tumours.
BibTeX
@inproceedings{lu2023ct,
  author = {Yang Lu and Haifeng Liu and Qi Liu and Siqi Wang and Zuhui Zhu and Jianguo Qiu and Wei Xing},
  title = {{CT-based radiomics with various classifiers for histological differentiation of parotid gland tumors}},
  booktitle = {Frontiers in Oncology},
  year = {2023},
  doi = {10.3389/fonc.2023.1118351},
}
Figure