AI-Assisted Ultrasound Review Before Thyroid Surgery

Number and proportion of enrolled participants who underwent at least one additional cytologic or pathologic examination, such as fine-needle aspiration or intraoperative frozen-section examination, after completion of conventional ultrasonography and AI-assisted review and before completion of surgery. Each participant is counted once regardless of the number of examinations. The decision remained clinician led; cervical lymph-node examinations were outside the AI system's task.

Number and proportion of enrolled participants with at least one documented change in the surgeon-led planned surgical approach or extent after conventional ultrasonography, AI-assisted review, and any supplementary pathological examination, when performed. Changes may involve thyroid resection or cervical lymph-node management. Each participant is counted once regardless of the number of changes. The AI system provided thyroid-nodule information only and did not assess cervical lymph nodes.

Number and percentage of evaluated thyroid nodules in each clinician-assigned C-TIRADS category during conventional ultrasonography. Nodule size, composition, echogenicity, margins, calcifications, and clinician-assessed cervical lymph-node findings were recorded as descriptive covariates and were not treated as separate outcome measures.

Number and percentage of evaluated thyroid nodules with a locked AI malignancy score >0.5. The AI system evaluated thyroid nodules only and did not assess cervical lymph nodes.

Among participants who underwent at least one additional cytologic or pathologic examination, record the cytologic or histopathologic result for each sampled thyroid nodule or cervical lymph node. Results are summarized at the examination level as nonmalignant, indeterminate, malignant, or metastatic, as applicable.

For each resected thyroid nodule linked unambiguously to the evaluated lesion, record the final diagnosis as benign, malignant, follicular tumor of uncertain malignant potential (FT-UMP), well-differentiated tumor of uncertain malignant potential (WDT-UMP), non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP), or another borderline diagnosis.

Using definitive postoperative histopathology as the reference standard, sensitivity is TP/(TP + FN), expressed as a percentage, for the locked AI classification (positive if score >0.5) and conventional C-TIRADS (positive if category 4a or higher). Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage. These lesion-level exclusions do not alter the 515-participant workflow cohort.

For each method, the false-negative rate is FN/(TP + FN), expressed as a percentage, among definitively malignant thyroid nodules. AI positivity is defined as score >0.5; conventional ultrasound positivity is C-TIRADS category 4a or higher. Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage.

For each method, positive predictive value is TP/(TP + FP), expressed as a percentage, among thyroid nodules classified as positive. AI positivity is defined as score >0.5; conventional ultrasound positivity is C-TIRADS category 4a or higher. Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage.

For each method, the Youden index is sensitivity + specificity - 1 using fixed binary thresholds. AI positivity is defined as score >0.5; conventional ultrasound positivity is C-TIRADS category 4a or higher. Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage.

Among cervical lymph nodes actually removed or sampled and pathologically examined, record the number and percentage with and without metastasis. No pathologic reference diagnosis is assigned to unsampled lymph nodes.

Using definitive postoperative histopathology as the reference standard, specificity is TN/(TN + FP), expressed as a percentage, for the locked AI classification (positive if score >0.5) and conventional C-TIRADS (positive if category 4a or higher). Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage.

Using definitive postoperative histopathology as the reference standard, negative predictive value is TN/(TN + FN), expressed as a percentage, for the locked AI classification (positive if score >0.5) and conventional C-TIRADS (positive if category 4a or higher). Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage.

Using definitive postoperative histopathology as the reference standard, accuracy is (TP + TN)/(TP + TN + FP + FN), expressed as a percentage, for the locked AI classification (positive if score >0.5) and conventional C-TIRADS (positive if category 4a or higher). Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage.

Using definitive postoperative histopathology as the reference standard, calculate the area under the receiver operating characteristic curve for the continuous locked AI malignancy score. Include only pathology-matched nodules with an unambiguous lesion-level link and a definitive benign or malignant diagnosis. Exclude FT-UMP, WDT-UMP, NIFTP, other borderline diagnoses, and nodules without unambiguous lesion-level linkage. No AUC is calculated for the single-threshold C-TIRADS classification.