Ultrasound deep learning features for predicting cervical lymph node metastasis in papillary thyroid carcinoma.

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Papillary thyroid cancer (PTC) management requires accurate assessment of cervical lymph node metastasis (CLNM).
• Current diagnostic methods for CLNM can be invasive or lack sufficient accuracy.
• Deep learning radiomics offers a potential noninvasive tool for CLNM detection.

Purpose of the Study:

• To develop and validate a clinical deep learning radiomics (DLR) model for determining CLNM status in PTC patients.
• To compare the performance of the DLR model against clinical models and experienced radiologists.

Main Methods:

• Retrospective enrollment of 205 PTC patients with preoperative thyroid ultrasonography (US) data.
• Development of three models: DLR, clinical, and a combined Cli-DLR model.
• Validation of models using a separate cohort (n=62) with analysis of AUC, precision-recall, and calibration.

Main Results:

• The combined Cli-DLR model demonstrated strong performance in the validation cohort with an AUC of 0.80.
• The Cli-DLR model showed excellent precision-recall and calibration.
• The Cli-DLR model outperformed both the clinical model and experienced radiologists in predicting CLNM.

Conclusions:

• A validated Cli-DLR model effectively assesses preoperative CLNM status in PTC patients.
• The model offers a noninvasive method for CLNM detection with favorable specificity and sensitivity.
• The Cli-DLR model shows potential to guide CLNM management and prevent overtreatment, pending further prospective validation.