Contrast-enhanced ultrasound-based energy response phenotyping in thyroid nodule ablation: identifying heterogeneous subgroups via latent class analysis

Introduction

Thermal ablation has emerged as an effective alternative to surgery for benign thyroid nodules, demonstrating significant volume reduction and symptom relief (1-3). However, long-term observations reveal substantial inter-patient heterogeneity, with regrowth rates of 10–35% at 3–5 years and unpredictable volume reduction rates (VRRs) (4,5). This variability reflects the biological complexity of nodule-energy interactions, where identical ablation protocols yield markedly different therapeutic responses. Current practice lacks reliable pre-procedural tools to stratify patients by anticipated efficacy, leading to suboptimal energy delivery and repeat interventions in resistant cases.

Existing research has focused on optimizing ablation techniques or identifying single imaging predictors, yet fails to capture the multidimensional nature of treatment response. Contrast-enhanced ultrasound (CEUS) can characterize nodule perfusion dynamics through quantitative parameters including peak intensity (PI), time to peak (TTP), and area under the curve (AUC), which reflect microvascular architecture and influence thermal energy distribution (6,7). However, these perfusion metrics have been studied in isolation rather than integrated with ablation parameters and longitudinal outcomes to define biologically meaningful patient subgroups.

Latent class analysis (LCA) offers an unsupervised learning approach to identify hidden patient subgroups based on multidimensional clinical data. This methodology has successfully revealed distinct phenotypes with differential treatment responses in acute respiratory distress syndrome (ARDS), hepatocellular carcinoma, and other heterogeneous diseases (8,9). By integrating pre-treatment perfusion signatures, energy delivery, and temporal volume reduction patterns, LCA can uncover clinically relevant energy response phenotypes that traditional univariate analyses miss. Such phenotypic stratification aligns with precision medicine principles, where treatment optimization depends on accurate patient classification using multiparametric biomarker profiles (10).

We employed LCA to identify distinct energy response phenotypes among patients undergoing thyroid nodule ablation, integrating quantitative CEUS parameters with ablation energy and longitudinal volume reduction trajectories. Subsequently, we developed a multinomial logistic regression model for pre-procedural phenotype prediction using baseline imaging and clinical variables. This approach addresses a critical clinical need: identifying patients who will achieve optimal response versus those requiring alternative energy strategies. By characterizing the biological underpinnings of treatment resistance through perfusion-based phenotyping, our findings provide a foundation for personalized ablation protocols that match energy delivery to individual nodule characteristics. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0756/rc) (11).

Methods

Study design and patient selection

This retrospective cohort study analyzed patients who underwent thermal ablation for benign thyroid nodules. The study protocol was approved by the Ethics Committee of The Central Hospital of Wuhan (No. 2025XXM-03-8337). Written informed consent for the ablation procedure was obtained from all patients. The requirement for additional study-specific consent was waived due to the retrospective nature of the analysis. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Inclusion criteria comprised: (I) benign thyroid nodules confirmed by fine-needle aspiration cytology; (II) complete pre-ablation CEUS examination; (III) documented ablation parameters (power, time, total energy); and (IV) longitudinal follow-up with volume measurements at 3, 6, and 12 months post-ablation. Exclusion criteria included: (I) malignant or indeterminate cytology; (II) incomplete CEUS data; (III) loss to follow-up before 12 months; and (IV) concurrent thyroid malignancy or previous thyroid surgery. Of 189 initially screened patients, 32 were excluded (indeterminate/malignant cytology n=8, incomplete CEUS data n=11, lost to follow-up n=13), yielding a final cohort of 157 patients (Figure S1).

Thermal ablation procedure

All ablations were performed using microwave ablation (MWA) with an ECO-100C MWA system equipped with 16-gauge cooled-shaft antennas. Ultrasound guidance was provided by a Mindray NUEWA R9S ultrasound system with an L14-3WU linear array transducer. The moving-shot technique was employed, wherein the antenna was systematically repositioned to achieve complete ablation coverage with real-time monitoring of the hyperechoic ablation zone.

Hydrodissection with normal saline was performed to maintain safe access routes and prevent thermal injury to adjacent structures. Injection sites were individualized based on nodule location: lateral to the trachea, anterior to the thyroid capsule, and/or posterior to the thyroid gland. Strict safety margins were maintained according to nodule anatomical position: ≥2 mm from the recurrent laryngeal nerve, trachea, esophagus, common carotid artery, internal jugular vein, superior laryngeal nerve, and parathyroid glands; ≥2 mm from the thyroid capsule when nodules abutted the lateral or posterior capsule. For isthmus nodules, bilateral recurrent laryngeal nerve trajectories were avoided with ≥2 mm clearance from the trachea. For upper pole nodules, the external branch of the superior laryngeal nerve was carefully avoided with 2 mm preservation of normal thyroid tissue at the apex. For mid-to-lower lateral lobe nodules (highest risk zone), a posterior “isolation zone” of ≥2 mm was strictly maintained to prevent direct contact between the ablation zone and the recurrent laryngeal nerve or parathyroid glands. When nodules were adjacent to major vessels, the ablation zone was kept ≥2 mm from the vessel wall using lower power and shorter duration settings.

Intraoperative saline injection during ablation served dual purposes: thermal protection and active heat dissipation from surrounding tissues. Post-ablation ice packs and additional hydrodissection (lateral to trachea, anterior and posterior to thyroid) were applied to reduce pain and facilitate tissue cooling. Immediately post-ablation, residual vascularity was assessed using CEUS; procedures were extended if significant peripheral perfusion persisted. All procedures were performed by two experienced interventional radiologists (each with >5 years and >500 thyroid ablation procedures). Patients received local anesthesia with 2% lidocaine and were monitored for 2 hours post-procedure before discharge.

Data collection and clinical assessment

Baseline data were systematically extracted from electronic medical records, including: (I) demographic characteristics (age, sex); (II) comorbidities (hypertension, diabetes mellitus, coronary artery disease); (III) nodule characteristics from conventional ultrasound (maximum diameter, volume calculated using the ellipsoid formula: π/6 × length × width × height, internal composition, presence of calcification); (IV) laboratory parameters obtained within one week pre-ablation [complete blood count, thyroid function tests including free triiodothyronine (T3), free thyroxine (T4), thyroid stimulating hormone (TSH), total T3, total T4, thyroglobulin antibody, thyroid peroxidase antibody, and thyroglobulin]; and (V) inflammatory markers calculated as neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR).

Ablation parameters (power settings, total ablation time, cumulative energy delivery) were recorded from procedure logs.

Post-ablation nodule volumes were measured at standardized follow-up intervals (3, 6, and 12 months) using the same ellipsoid formula, with VRR calculated as: VRR (%) = [(baseline volume − follow-up volume)/baseline volume] × 100.

Missing data were handled using multivariate imputation by chained equations (MICE) with predictive mean matching (PMM) method (50 iterations, 5 imputed datasets). Auxiliary variables included nodule volume, CEUS parameters, and ablation energy. Convergence was verified by trace plot inspection. Pooled estimates were obtained using Rubin’s rules.

CEUS examination and quantitative analysis

Pre-ablation CEUS examinations were performed using a diagnostic ultrasound system with a high-frequency linear transducer (5–12 MHz). Following intravenous bolus injection of 2.4 mL sulfur hexafluoride microbubbles (SonoVue, Bracco, Milan, Italy), real-time perfusion imaging was acquired for 3 minutes. Quantitative CEUS parameters were extracted using dedicated software (VueBox, Bracco): PI (dB), TTP (seconds), and AUC. Enhancement patterns were categorized as non-enhancement, hypo-enhancement, iso-enhancement, or hyper-enhancement relative to adjacent thyroid parenchyma. All measurements were performed by two experienced radiologists blinded to clinical outcomes, with inter-observer agreement assessed using intraclass correlation coefficients (ICC >0.85 for all parameters). Representative pre-ablation and post-ablation ultrasound images demonstrating volume reduction are presented in Figure 1.

Figure 1 Representative ultrasound images of a thyroid nodule before and 12 months after thermal ablation. Pre-ablation imaging (A) demonstrates the target nodule prior to treatment. Follow-up imaging at 12 months (B) shows complete resolution of the nodule, corresponding to a VRR of 100%, representative of the class 3 optimal-response phenotype. Class 3, optimal-response. VRR, volume reduction rate.

LCA for phenotype identification

Unsupervised LCA was employed to identify distinct energy response phenotypes based on integrated CEUS perfusion metrics (PI, AUC, TTP), ablation energy parameters, and longitudinal VRR trajectories at 3, 6, and 12 months. Continuous variables were categorized into tertiles (low/medium/high) to meet LCA assumptions. Models with 2–4 latent classes were fitted using the poLCA package in R, with 20 random starts per model to avoid local maxima. Model selection balanced statistical criteria [Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), entropy >0.80] with clinical interpretability. The optimal model was selected based on parsimony, classification quality (entropy), and biological plausibility of identified phenotypes.

Discriminant model development and validation

A multinomial logistic regression model was constructed to predict phenotype membership using pre-ablation variables: CEUS parameters (PI, AUC, TTP), nodule characteristics (volume, maximum diameter, internal composition, calcification), enhancement pattern, and baseline laboratory indices (complete blood count, inflammatory markers). The dataset was randomly partitioned into training (70%) and testing (30%) sets using stratified sampling to maintain class proportions. Continuous predictors were standardized [mean =0, standard deviation (SD) =1] prior to modeling. Stepwise bidirectional variable selection was performed using the AIC. Model performance was evaluated using overall accuracy, Cohen’s kappa coefficient, and class-specific metrics (sensitivity, specificity, positive/negative predictive values, F1-score). Discriminative capacity was assessed via one-vs-rest receiver operating characteristic (ROC) analysis with AUC calculation. Ten-fold cross-validation was conducted on the training set to assess model stability and generalizability. To address potential model instability suggested by large coefficients, we additionally performed penalized multinomial logistic regression using elastic net regularization (alpha =0.5, lambda selected by 10-fold cross-validation) and compared results with the standard model.

Statistical analysis

Continuous variables were assessed for normality using Shapiro-Wilk tests. Normally distributed data were presented as mean ± SD and compared using one-way analysis of variance (ANOVA); non-normally distributed data were presented as median [interquartile range (IQR)] and compared using Kruskal-Wallis tests. Categorical variables were expressed as frequencies (percentages) and compared using chi-square or Fisher’s exact tests. Variable importance in the discriminant model was quantified by mean absolute coefficient values across all phenotype comparisons. Odds ratios (ORs) with 95% confidence intervals (CIs) were calculated for top predictors using class 1 as the reference category. Imputation was performed on the complete cohort prior to the training/testing partition to avoid information leakage. Ten-fold cross-validation was conducted exclusively within the training set (n=110). All statistical analyses were performed using R version 4.3.0 (R Foundation for Statistical Computing), with two-sided P<0.05 considered statistically significant.

Results

Patient characteristics and latent class identification

A total of 157 patients (84.7% female; median age 47 years) who underwent thermal ablation for thyroid nodules were included in this retrospective analysis. The median nodule volume was 310.45 mm³ (IQR: 113.10–3,655.49 mm³), with predominantly solid composition (59.9%). Pre-ablation CEUS parameters demonstrated considerable heterogeneity including median PI 26.67 dB (IQR: 23.46–30.36 dB), TTP 22.59 s (IQR: 17.93–26.88 s), and AUC 1,466.24 (IQR: 1,279.24–1,657.00). Median total ablation energy was 1,710.00 J (IQR: 1,000.00–6,480.00 J), with VRR progressively increasing from 62.58% at 3 months to 92.62% at 12 months (Table 1).

LCA models with 2, 3, and 4 classes were fitted. Model fit statistics are summarized in Table 2. Although the 2-class model demonstrated the lowest BIC (2,443.36 vs. 2,464.08 for 3-class), the 3-class solution was selected based on: (I) superior clinical interpretability with three distinct energy-response patterns; (II) adequate entropy (0.833) indicating good classification certainty; and (III) meaningful treatment stratification potential. The 4-class model showed further BIC increase (2,502.86) without additional clinical utility. This exploratory model selection prioritized clinical interpretability alongside statistical parsimony. The three identified phenotypes were clinically labeled based on their temporal response patterns and energy-outcome relationships: class 1 as “delayed-response” (n=56, 35.7%), class 2 as “energy-resistant” (n=33, 21.0%), and class 3 as “optimal-response” (n=68, 43.3%) (Figure 2).

Figure 2 Longitudinal analysis of VRR trajectories across three latent classes. Class 1, delayed-response; class 2, energy-resistant; class 3, optimal-response. VRR, volume reduction rate.

Phenotypic characterization and CEUS perfusion signatures

The three latent classes exhibited markedly divergent CEUS perfusion profiles, energy requirements, and temporal response patterns (Table 3).

Class 1 (“delayed-response phenotype”) demonstrated the highest pre-ablation PI (30.74 dB, P<0.001) and most prolonged TTP (P<0.001). This phenotype exhibited modest early-phase volume reduction at 3 months (VRR 60.60%) and 6 months (VRR 75.53%), followed by substantial delayed improvement at 12 months (VRR 89.86%), suggesting gradual thermal injury propagation despite initial high perfusion intensity. Class 2 (“energy-resistant phenotype”) presented with intermediate PI (24.63 dB) and the shortest TTP (18.27 s, P<0.001), indicating rapid contrast washout. Despite requiring substantially higher energy delivery (median 7,595.00 J, P<0.001), this class consistently demonstrated the poorest treatment outcomes including VRR of 60.65% at 3 months, 75.31% at 6 months, and 86.56% at 12 months (P<0.001). Class 3 (“optimal-response phenotype”) displayed moderate PI (24.66 dB) and TTP (20.48 s). This phenotype demonstrated superior treatment sensitivity, achieving the highest VRR at all follow-up intervals with 73.45% at 3 months, 90.26% at 6 months, and 97.65% at 12 months (P<0.001).

Baseline characteristics including age (P=0.125) and inflammatory markers (NLR P=0.541, PLR P=0.836) did not differ significantly across phenotypes. Nodule maximum diameter (P<0.001) and volume (P<0.001) differed significantly, with Class 2 exhibiting larger nodules consistent with higher energy requirements.

Discriminant model performance and predictive capacity

All 157 patients had complete data for candidate predictors and were included in discriminant model development. The cohort was randomly divided into training (n=110, 70%) and testing (n=47, 30%) sets with stratified sampling to maintain class proportions. A multinomial logistic regression model incorporating nine predictors [pre-ablation PI, AUC, TTP, nodule volume, white blood cell (WBC), platelet count, neutrophil count, NLR, PLR] achieved moderate discrimination.

In the training set, the model demonstrated an overall accuracy of 71.56% (kappa =0.545), with class-specific performance showing class 1 sensitivity 84.2% (specificity 85.9%), class 2 sensitivity 39.1% (specificity 94.2%), and class 3 sensitivity 77.1% (specificity 73.8%). In the independent testing cohort (n=47), overall accuracy was 61.70% (kappa =0.399), with class 1 sensitivity 82.4% (specificity 70.0%), class 2 sensitivity 50.0% (specificity 94.6%), and class 3 sensitivity 50.0% (specificity 74.1%) (Tables S1,S2).

ROC analysis using the one-vs-rest approach (Figure 3A) demonstrated class-specific AUC values: class 1 (delayed-response) AUC =0.878 (95% CI: 0.771–0.986), class 2 (energy-resistant) AUC =0.916 (95% CI: 0.837–0.995), and class 3 (optimal-response) AUC =0.726 (95% CI: 0.580–0.872), with mean AUC =0.840. The heatmap of classification probabilities further illustrated the model’s discriminative performance across the three phenotypes (Figure 3B). Ten-fold cross-validation in the full cohort yielded mean accuracy 59.73%±15.93% (kappa =0.328±0.209), indicating preliminary discriminative capacity with substantial variability (Figure 3C, Figure S2).

Figure 3 Latent class analysis results and discriminant model performance. (A) ROC curves for three energy response phenotypes. (B) Heatmap visualization of classification probabilities. (C) Bar chart of 10-fold cross-validation accuracy. Class 1, delayed-response; class 2, energy-resistant; class 3, optimal-response. AUC, area under the curve; CV, cross-validation; ROC, receiver operating characteristic.

Variable importance analysis (Figure 4A) identified pre-ablation TTP as the dominant predictor (relative importance: 100%), followed by PI (56.2%), neutrophil count (43.3%), WBC count (22.1%), and NLR (10.1%). Coefficient heatmap visualization (Figure 4B) revealed distinct predictor-phenotype associations: class 2 (vs. class 1) exhibited strong negative associations with PI (OR =0.094, P<0.001) and TTP (OR =0.106, P<0.001), alongside positive associations with WBC (OR =1.878, P=0.035) and neutrophil count (OR =2.562, P=0.033), consistent with the energy-resistant profile characterized by reduced perfusion and systemic inflammatory markers. Class 3 (vs. class 1) demonstrated a strong negative association with TTP (OR =0.122, P<0.001) and positive association with neutrophil count (OR =2.561, P=0.005) (Table S3).

Figure 4 Variable importance and model coefficients analysis. (A) Variable importance ranking based on mean absolute coefficient values across all phenotype comparisons, with relative importance scaled to 0–100%. (B) Heatmap of model coefficients for each predictor-phenotype combination. Class 1 serves as the reference category (all coefficients =0). Color gradient: blue (negative coefficients) to red (positive coefficients). (C) Forest plot showing OR with 95% confidence intervals for the top 5 most important variables. Points represent OR estimates; horizontal lines indicate 95% CI; vertical dashed line at OR =1 represents no effect. Pre-ablation CEUS parameters (PI, TTP) emerge as the strongest predictors of energy response phenotype. *, P<0.05; **, P<0.01. AUC, area under the curve; CEUS, contrast-enhanced ultrasound; CI, confidence interval; NLR, neutrophil-to-lymphocyte ratio; OR, odds ratio; PI, peak intensity; PLR, platelet-to-lymphocyte ratio; TTP, time to peak; WBC, white blood cell.

Forest plot analysis (Figure 4C) confirmed TTP and PI exerted the strongest discriminative effects across phenotype comparisons, with ORs indicating that lower TTP and PI values substantially increased the probability of class 2 membership relative to class 1, reflecting the energy-resistant phenotype’s characteristic rapid washout and reduced peak perfusion.

Discussion

This study represents the first application of LCA to identify energy response phenotypes in thyroid nodule thermal ablation, addressing a critical clinical problem where long-term regrowth rates range from 10–35% (4,10). We identified three distinct phenotypes characterized by divergent CEUS perfusion properties: a delayed-response phenotype (class 1) with high PI and prolonged TTP, an energy-resistant phenotype (class 2) requiring 7.6-fold higher energy yet achieving suboptimal outcomes, and an optimal-response phenotype (class 3) with balanced perfusion achieving 97% volume reduction at 12 months. A preliminary multinomial predictive model incorporating pre-ablation CEUS parameters and clinical variables demonstrated moderate discrimination (mean AUC =0.840, testing accuracy =61.70%), suggesting potential utility for prospective phenotype identification pending external validation.

Our findings address substantial heterogeneity in thyroid ablation outcomes, with regrowth representing the primary clinical challenge (1). Previous investigations focused predominantly on technical optimization or single imaging predictors. While CEUS has enhanced malignancy risk stratification in thyroid nodules (10), and quantitative parameters predict diagnostic accuracy (12,13), these studies primarily addressed characterization rather than treatment prediction. European guidelines established CEUS for assessing ablation completeness (10), yet integration with energy delivery and longitudinal outcomes remains limited. Our phenotyping approach extends this foundation by linking pre-treatment perfusion signatures to differential treatment responses.

Building upon these observations, the LCA methodology draws conceptual support from successful applications in heterogeneous disease states. Sinha and colleagues identified two ARDS phenotypes displaying differential treatment responses to fluid management and corticosteroids, with the hyperinflammatory phenotype demonstrating improved survival with specific therapies (14-16). Similar phenotypic classifications in coronavirus disease 2019 (COVID-19) ARDS (17), chronic obstructive pulmonary disease (18), and cardiovascular conditions (19) confirmed reproducibility and prognostic significance (20). These precedents validate our methodological approach while highlighting thyroid ablation’s unique challenge of that phenotypes emerge not from systemic inflammatory states but from local tissue perfusion heterogeneity determining thermal energy propagation.

Based on these findings, we explored the biological mechanisms explaining phenotypic divergence. Class 2’s energy-resistant phenotype, characterized by large volume, reduced PI, and rapid washout, likely reflects inadequate microvascular density and arteriovenous shunting. This hemodynamic profile fundamentally impedes thermal energy distribution through multiple mechanisms: insufficient capillary density limits heat deposition, creating undertreated peripheral zones; arteriovenous shunting increases convective heat loss preventing tissue temperatures from reaching coagulation thresholds; and large baseline volume exceeds the effective treatment zone achievable with standard protocols. The requirement for 7.6-fold higher energy despite persistent suboptimal outcomes (12-month VRR 86.56% vs. 97.65% for class 3) suggests fundamental limitations in achieving thermal coagulation despite aggressive escalation. Notably, despite receiving the highest absolute energy, class 2’s energy density (3,489.54 J/mL) was substantially lower than class 3 (8,914.42 J/mL, P=0.038), confirming that large nodule volume diluted the thermal dose and contributed to treatment resistance.

This mirrors observations in hepatocellular carcinoma ablation, where hypervascular tumors demonstrate higher recurrence rates due to heat sink effects from peritumoral vessels (21). Conversely, class 3’s optimal-response phenotype benefits from moderate perfusion maintaining tissue viability without excessive heat dissipation, smaller baseline volume enabling complete energy coverage, and balanced vascular architecture allowing adequate thermal coagulation without premature washout. This is reflected in the highest energy density among all phenotypes (8,914.42 J/mL), indicating concentrated thermal delivery per unit volume. Class 1’s delayed-response pattern suggests compensatory mechanisms: high initial perfusion may trigger progressive vascular injury cascades extending beyond the immediate thermal zone, with delayed thrombosis and ischemic propagation contributing to gradual volume reduction over 12 months.

Pre-ablation PI emerged as the dominant discriminator (100% relative importance), with PI contributing 56.2% relative importance, underscoring microvascular perfusion dynamics’ central role in determining treatment response. This aligns with evidence demonstrating that perfusion imaging predicts tumor interventional treatment response across multiple modalities (22,23). The contrast kinetics reflected in TTP and PI provide complementary information: TTP captures vascular architectural complexity and flow velocity, while PI quantifies microvascular density and permeability. Their combined contribution corroborates studies showing that multiparametric perfusion assessment outperforms single-parameter approaches in predicting therapeutic outcomes (24).

These mechanistic insights provide a foundation for clinical translation. The pre-ablation predictive model addresses a critical need: prospectively identifying patients likely to achieve suboptimal outcomes before irreversible energy delivery. As thermal ablation gains acceptance as a surgical alternative (25), the ability to stratify treatment candidates becomes paramount. For class 2 patients, the model’s discrimination (AUC =0.916, testing sensitivity 40.0%) enables evidence-based counseling regarding alternative strategies: surgical referral for definitive resection, hybrid approaches combining cytoreductive debulking with subsequent ablation, or modified protocols incorporating advanced techniques such as marginal venous ablation or hydrodissection to minimize heat sink effects. Additionally, closer surveillance intervals permit earlier regrowth detection when repeat ablation remains feasible.

However, several limitations warrant acknowledgment. First, the single-center retrospective design introduces potential selection bias. Multi-center prospective validation is essential to confirm phenotype generalizability across diverse practice settings. Second, the 12-month follow-up precludes evaluation of long-term regrowth patterns; extended surveillance (3–5 years) is necessary to determine whether phenotypic classification predicts durable efficacy versus delayed recurrence. Third, the model incorporated only pre-ablation clinical and imaging variables, excluding intra-procedural parameters that might refine predictions. Fourth, the predictive model’s moderate testing accuracy (61.70%) and substantial cross-validation variability (59.73%±15.93%) indicate preliminary discriminative capacity requiring further refinement. The modest performance for class 3 discrimination (AUC =0.726) suggests additional predictors may be needed to fully capture optimal-response phenotype characteristics. Fifth, while we propose biological mechanisms linking perfusion phenotypes to treatment response, direct validation through histopathological correlation was not feasible in this clinical cohort.

Future research should pursue prospective randomized trials comparing phenotype-matched treatment strategies versus standard approaches to demonstrate clinical utility. Such trials could test whether intensive ablation protocols (higher energy density, extended ablation times, multi-session approaches) improve class 2 outcomes or whether surgical referral provides superior benefit. Integrating advanced imaging modalities—such as shear wave elastography to assess tissue stiffness or superb microvascular imaging to visualize microvascular architecture—could enhance phenotype characterization and prediction accuracy. Extending this phenotyping framework to other thermal ablation indications (hepatic tumors, renal masses, bone metastases) could broaden clinical applicability and test whether perfusion-based phenotyping represents a generalizable principle across organ systems. Finally, economic analyses quantifying cost-effectiveness of phenotype-guided personalized strategies versus empirical treatment represent critical considerations for healthcare system adoption and reimbursement policy.

Conclusions

This exploratory study establishes CEUS perfusion phenotypes as potentially meaningful classifiers for predicting thermal ablation response in thyroid nodules. The identification of three distinct energy response phenotypes, characterized by divergent perfusion dynamics, energy requirements, and temporal response patterns, provides a preliminary framework for precision medicine approaches in thyroid interventional radiology. Pre-procedural phenotype prediction using CEUS parameters and clinical variables demonstrated moderate discrimination (mean AUC =0.840), representing an initial step toward individualized treatment planning pending external validation. While methodological limitations and modest predictive accuracy necessitate cautious interpretation, this phenotyping approach advances beyond empiric protocols toward data-driven strategies that may ultimately reduce treatment failures and optimize patient selection for thermal ablation versus alternative management.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0756/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0756/dss

Funding: This work was supported by the Wuhan Municipal Natural Science Foundation—Key Special Project for Clinical Research of Municipal Medical Institutions (grant No. 2026020301040126).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0756/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study protocol was approved by the Ethics Committee of The Central Hospital of Wuhan (No. 2025XXM-03-8337). Written informed consent for the ablation procedure was obtained from all patients. The requirement for additional study-specific consent was waived due to the retrospective nature of the analysis. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Cho SJ, Baek JH, Chung SR, Choi YJ, Lee JH. Long-Term Results of Thermal Ablation of Benign Thyroid Nodules: A Systematic Review and Meta-Analysis. Endocrinol Metab (Seoul) 2020;35:339-50. [Crossref] [PubMed]
2. Yildirim G, Karakas HM. Uncooled Microwave Ablation as a Treatment Option to Preserve Thyroid Function in Patients with Benign Thyroid Nodules. J Belg Soc Radiol 2022;106:50. [Crossref] [PubMed]
3. Sinha P, Furfaro D, Cummings MJ, Abrams D, Delucchi K, Maddali MV, et al. Latent Class Analysis Reveals COVID-19-related Acute Respiratory Distress Syndrome Subgroups with Differential Responses to Corticosteroids. Am J Respir Crit Care Med 2021;204:1274-85. [Crossref] [PubMed]
4. Sim JS, Baek JH. Long-Term Outcomes Following Thermal Ablation of Benign Thyroid Nodules as an Alternative to Surgery: The Importance of Controlling Regrowth. Endocrinol Metab (Seoul) 2019;34:117-23. [Crossref] [PubMed]
5. Jeong SY, Baek JH. Long-term clinical outcomes of thermal ablation for benign thyroid nodules and unresolved issues: a comprehensive systematic review. Radiol Med 2025;130:111-20. [Crossref] [PubMed]
6. Chen F, Han H, Wan P, Chen L, Kong W, Liao H, Wen B, Liu C, Zhang D. Do as Sonographers Think: Contrast-Enhanced Ultrasound for Thyroid Nodules Diagnosis via Microvascular Infiltrative Awareness. IEEE Trans Med Imaging 2024;43:3881-94. [Crossref] [PubMed]
7. Fan J, Tao L, Zhan W, Li W, Kuang L, Zhao Y, Zhou W. Diagnostic value of qualitative and quantitative parameters of contrast-enhanced ultrasound for differentiating differentiated thyroid carcinomas from benign nodules. Front Endocrinol (Lausanne) 2023;14:1240615. [Crossref] [PubMed]
8. Sinha P, Calfee CS, Delucchi KL. Practitioner's Guide to Latent Class Analysis: Methodological Considerations and Common Pitfalls. Crit Care Med 2021;49:e63-79. [Crossref] [PubMed]
9. Chen S, Guo A, Lu L, Lin S, Hu X, Zhu L, Chen X. Latent Class Analysis of Subphenotypes in Intermediate-Stage Hepatocellular Carcinoma after Transarterial Chemoembolization. J Cancer 2022;13:3318-25. [Crossref] [PubMed]
10. Radzina M, Ratniece M, Putrins DS, Saule L, Cantisani V. Performance of Contrast-Enhanced Ultrasound in Thyroid Nodules: Review of Current State and Future Perspectives. Cancers (Basel) 2021;13:5469. [Crossref] [PubMed]
11. Moons KG, Altman DG, Reitsma JB, Ioannidis JP, Macaskill P, Steyerberg EW, Vickers AJ, Ransohoff DF, Collins GS. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med 2015;162:W1-73. [Crossref] [PubMed]
12. Liu Y, Liu H, Zhan J, Chai Q, Zhu J, Ding S, Chen L. Contrast-Enhanced Ultrasound for Diagnosing Thyroid Nodules With Indeterminate Cytology: A Retrospective Study. Clin Endocrinol (Oxf) 2025;102:223-31. [Crossref] [PubMed]
13. Mauri G, Pacella CM, Papini E, Solbiati L, Goldberg SN, Ahmed M, Sconfienza LM. Image-Guided Thyroid Ablation: Proposal for Standardization of Terminology and Reporting Criteria. Thyroid 2019;29:611-8. [Crossref] [PubMed]
14. Sinha P, Delucchi KL, McAuley DF, O'Kane CM, Matthay MA, Calfee CS. Development and validation of parsimonious algorithms to classify acute respiratory distress syndrome phenotypes: a secondary analysis of randomised controlled trials. Lancet Respir Med 2020;8:247-57. [Crossref] [PubMed]
15. Sinha P, Churpek MM, Calfee CS. Machine Learning Classifier Models Can Identify Acute Respiratory Distress Syndrome Phenotypes Using Readily Available Clinical Data. Am J Respir Crit Care Med 2020;202:996-1004. [Crossref] [PubMed]
16. Calfee CS, Delucchi KL, Sinha P, Matthay MA, Hackett J, Shankar-Hari M, McDowell C, Laffey JG, O'Kane CM, McAuley DF. Acute respiratory distress syndrome subphenotypes and differential response to simvastatin: secondary analysis of a randomised controlled trial. Lancet Respir Med 2018;6:691-8. [Crossref] [PubMed]
17. Sinha P, Matthay MA, Calfee CS. Is a "Cytokine Storm" Relevant to COVID-19? JAMA Intern Med 2020;180:1152-4. [Crossref] [PubMed]
18. Burgel PR, Paillasseur JL, Janssens W, Piquet J, Ter Riet G, Garcia-Aymerich J, et al. A simple algorithm for the identification of clinical COPD phenotypes. Eur Respir J 2017;50:1701034. [Crossref] [PubMed]
19. Inohara T, Shrader P, Pieper K, Blanco RG, Thomas L, Singer DE, Freeman JV, Allen LA, Fonarow GC, Gersh B, Ezekowitz MD, Kowey PR, Reiffel JA, Naccarelli GV, Chan PS, Steinberg BA, Peterson ED, Piccini JP. Association of of Atrial Fibrillation Clinical Phenotypes With Treatment Patterns and Outcomes: A Multicenter Registry Study. JAMA Cardiol 2018;3:54-63. [Crossref] [PubMed]
20. Jameson JL, Longo DL. Precision medicine--personalized, problematic, and promising. N Engl J Med 2015;372:2229-34. [Crossref] [PubMed]
21. Lu DS, Raman SS, Limanond P, Aziz D, Economou J, Busuttil R, Sayre J. Influence of large peritumoral vessels on outcome of radiofrequency ablation of liver tumors. J Vasc Interv Radiol 2003;14:1267-74. [Crossref] [PubMed]
22. El Kaffas A, Hoogi A, Zhou J, Durot I, Wang H, Rosenberg J, Tseng A, Sagreiya H, Akhbardeh A, Rubin DL, Kamaya A, Hristov D, Willmann JK. Spatial Characterization of Tumor Perfusion Properties from 3D DCE-US Perfusion Maps are Early Predictors of Cancer Treatment Response. Sci Rep 2020;10:6996. [Crossref] [PubMed]
23. Goh V, Halligan S, Hugill JA, Bassett P, Bartram CI. Quantitative assessment of colorectal cancer perfusion using MDCT: inter- and intraobserver agreement. AJR Am J Roentgenol 2005;185:225-31. [Crossref] [PubMed]
24. Jiang T, Kambadakone A, Kulkarni NM, Zhu AX, Sahani DV. Monitoring response to antiangiogenic treatment and predicting outcomes in advanced hepatocellular carcinoma using image biomarkers, CT perfusion, tumor density, and tumor size (RECIST). Invest Radiol 2012;47:11-7. [Crossref] [PubMed]
25. Papini E, Monpeyssen H, Frasoldati A, Hegedüs L. 2020 European Thyroid Association Clinical Practice Guideline for the Use of Image-Guided Ablation in Benign Thyroid Nodules. Eur Thyroid J 2020;9:172-85. [Crossref] [PubMed]