Application of AmCAD‐UT in Active Surveillance of Papillary Thyroid Microcarcinoma and Suspicious Subcentimeter Thyroid Nodules: A Prospective Agreement Study

Bland–Altman plots for agreement between radiologist and AmCAD‑UT measurements. Artificial intelligence (AI)–assisted computer-aided diagnosis (CAD) has shown promising diagnostic performance for thyroid ultrasound, but its reliability for longitudinal active surveillance (AS) workflows remains insufficiently characterized [1-3]. Interobserver variability in ultrasound interpretation further motivates tools that can standardize measurements and reporting [4]. We evaluated whether AmCAD-UT (Artificial Intelligence–based Computer-Aided Diagnosis for Ultrasound Thyroid) provides reproducible nodule measurements and consistent risk stratification during AS of papillary thyroid microcarcinoma (PTMC) and suspicious subcentimeter thyroid micro-nodules (STMN). From June 2022 to August 2024, we consecutively enrolled 100 adults managed with AS at Ruijin Hospital (50 cytology-confirmed PTMC; 50 STMN without prior cytology). Eligibility criteria included maximal nodule diameter < 10 mm, no extrathyroidal extension, no suspicious cervical lymph nodes, and adherence to the institutional AS protocol [5-7]. Each participant underwent 2 ultrasound examinations (baseline and follow-up separated by ≥ 6 months; total 200 examinations). Radiologist measurements and categorizations were compared with AmCAD-UT outputs. Agreement in three-diameter measurements was assessed using paired testing and Bland–Altman analyses (mean bias and 95% limits of agreement). Agreement in TIRADS (Thyroid Imaging Reporting and Data System) categories and key sonographic features (orientation, margin, calcification) was assessed using weighted κ with 95% confidence intervals (2-sided α = 0.05). Two TIRADS systems were applied: the American College of Radiology TIRADS (ACR TIRADS) and the Chinese TIRADS (C-TIRADS). ACR TIRADS is a point-based system that assigns weighted scores to sonographic features including composition, echogenicity, shape, margin, and echogenic foci to stratify malignancy risk, whereas C-TIRADS uses a simplified pattern-based classification derived from multicenter Chinese cohorts, emphasizing key suspicious features with a more streamlined scoring approach [8]. The study was approved by the institutional review board, and all participants provided written informed consent. Analyses were performed using SPSS version 26 and R. Median follow-up was 15.5 months (IQR, 7.8–23.3) for PTMC and 12.0 months (IQR, 8.5–20.0) for STMN. Median age was 40.5 years (IQR, 34.0–48.0) for PTMC (female, 33/50 [66.0%]) and 41.5 years (IQR, 36.0–55.0) for STMN (female, 41/50 [82.0%]). Microcalcifications were observed in 41.0% of PTMC and 33.0% of STMN by radiologist assessment. AmCAD-UT and radiologist measurements were highly consistent with no statistically significant differences in mean transverse, anteroposterior, or longitudinal diameters for either cohort (Table 1). Bland–Altman plots (Figure 1) demonstrated minimal mean bias with narrow limits of agreement across all 3 diameters. Bland–Altman plots for agreement between radiologist and AmCAD‑UT measurements. (Panels A–C): PTMC (anteroposterior, transverse/horizontal, longitudinal/vertical). (Panels D–F): STMN (anteroposterior, transverse/horizontal, longitudinal/vertical). Central dashed line indicates mean difference; upper and lower dashed lines indicate 95% limits of agreement. In contrast, agreement in TIRADS categorization was limited. For PTMC, ACR TIRADS agreement was fair (κ = 0.321 at baseline; κ = 0.222 at follow-up), whereas C-TIRADS agreement was poor (κ = 0.008–0.134; p > 0.05). For STMN, agreement was fair for both ACR TIRADS and C-TIRADS (κ = 0.272–0.317 and κ = 0.277–0.283, respectively). Across cohorts and grading systems, AmCAD-UT showed a downgrading tendency (ACR downgraded 24%–34%; C-TIRADS downgraded 24%–28%) (Table 1). Feature-level agreement was higher for PTMC than STMN: orientation (κ 0.669–0.737 vs. 0.430–0.480), margins (κ 0.729–0.851 vs. 0.071–0.324), and calcification (κ 0.470–0.703 vs. 0.381–0.561). In this prospective AS cohort of subcentimeter thyroid lesions, AmCAD-UT produced reproducible size measurements comparable with expert radiologist assessment, supporting its potential role in standardizing longitudinal AS measurements where small absolute changes may trigger management escalation [5-7]. However, AI-derived TIRADS classifications based on static images showed only fair agreement with systematic downgrading. This pattern may reflect the loss of real-time sonographic context (e.g., dynamic evaluation of margins, echogenic foci, and artifacts) and/or threshold differences between the algorithm and clinician judgment; therefore, risk-stratification outputs should be used as an adjunct rather than a substitute for expert interpretation. Limitations include the single-center design, modest sample size, reliance on static images without real-time scanning context, and absence of dedicated cervical lymph node evaluation. Multicenter validation incorporating real-time imaging and lymph node assessment is warranted. Clinically, our findings support the use of AI primarily for consistent longitudinal measurement tracking in active surveillance, rather than for independent TI-RADS risk categorization. Yuxin Ge: writing – original draft, data curation, formal analysis. Bin Zheng: methodology, data curation. Yonggang He: conceptualization, methodology, supervision, funding acquisition, writing – review and editing. This research is supported by a clinical research grant of the Shanghai Municipal Health Commission (Grant 202040019). The authors declare no conflicts of interest. The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.