Enhanced Ultrasound Radiomics Using Super-Resolution Techniques for Predicting Nondiagnostic Thyroid Nodules
Overview
Revise to specify the comparative advantages of GAN over traditional ultrasound evaluations.
Background
Thyroid nodules are prevalent and pose a challenge in clinical endocrinology, particularly in ruling out malignancy. Nondiagnostic results from FNA, categorized as Bethesda I, occur in approximately 15% of cases, leading to increased patient anxiety and healthcare costs. Enhancing the predictive capabilities for these nodules is crucial for improving patient management and reducing unnecessary procedures.
Data Highlights
Model
AUC (Training)
AUC (Internal Testing)
AUC (Validation)
SR-RF
0.808
0.733
0.7435
NR-RF
0.672
0.596
N/A
Key Findings
The SR-RF model outperformed the NR-RF model in predicting nondiagnostic thyroid nodules.
Post hoc calibration improved the Brier score, indicating enhanced probability reliability.
False-positive results were often associated with cystic nodules and acoustic artifacts.
False-negative results were linked to isoechoic solid nodules with indistinct margins.
No significant impact of operator experience on model performance was observed.
Clinical Implications
The integration of GAN-based super-resolution radiomics into clinical practice could enhance the pre-FNA assessment of thyroid nodules, potentially reducing the number of unnecessary repeat aspirations. This approach may lead to more accurate risk stratification and improved patient outcomes.
Conclusion
The study highlights the potential of advanced ultrasound techniques to improve the diagnostic accuracy for nondiagnostic thyroid nodules, suggesting a shift towards more reliable preoperative assessments.
