Comparison of the diagnostic effectiveness of ultrasound imaging coupled with three mathematical models for discriminating thyroid nodules

Acta Radiol. 2024 May;65(5):441-448. doi: 10.1177/02841851231221912. Epub 2024 Jan 17.

Abstract

Background: The overlapping nature of thyroid lesions visualized on ultrasound (US) images could result in misdiagnosis and missed diagnoses in clinical practice.

Purpose: To compare the diagnostic effectiveness of US coupled with three mathematical models, namely logistic regression (Logistics), partial least-squares discriminant analysis (PLS-DA), and support vector machine (SVM), in discriminating between malignant and benign thyroid nodules.

Material and methods: A total of 588 thyroid nodules (287 benign and 301 malignant) were collected, among which 80% were utilized for constructing the mathematical models and the remaining 20% were used for internal validation. In addition, an external validation cohort comprising 160 nodules (80 benign and 80 malignant) was employed to validate the accuracy of these mathematical models.

Results: Our study demonstrated that all three models exhibited effective predictive capabilities for distinguishing between benign and malignant nodules, whose diagnostic effectiveness surpassed that of the TI-RADS classification, particularly in terms of true negative diagnoses. SVM achieved a higher diagnostic rate for malignant thyroid nodules (93.8%) compared to Logistics (91.5%) and PLS-DA (91.6%). PLS-DA exhibited higher diagnostic rates for benign thyroid nodules (91.9%) compared to Logistics (86.7%) and SVM (88.7%). Both the area under the receiver operating characteristic curve (AUC) values of PLS-DA (0.917) and SVM (0.913) were higher than that of Logistics (0.891).

Conclusion: Our findings indicate that SVM had significantly higher rates of true positive diagnoses and PLS-DA exhibited significantly higher rates of true negative diagnoses. All three models outperformed the TI-RADS classification in discriminating between malignant and benign thyroid nodules.

Keywords: Ultrasound; logistic regression; mathematical model; partial least squares discriminant analysis; support vector machine; thyroid nodule.

Publication types

  • Comparative Study

MeSH terms

  • Adolescent
  • Adult
  • Aged
  • Diagnosis, Differential
  • Discriminant Analysis
  • Female
  • Humans
  • Least-Squares Analysis
  • Logistic Models
  • Male
  • Middle Aged
  • Models, Theoretical
  • Reproducibility of Results
  • Retrospective Studies
  • Sensitivity and Specificity
  • Support Vector Machine
  • Thyroid Gland / diagnostic imaging
  • Thyroid Nodule* / diagnostic imaging
  • Ultrasonography* / methods
  • Young Adult