Predicting Extranodal Extension with Preoperative Contrast-enhanced CT in Patients with Oropharyngeal Squamous Cell Carcinoma

Abstract

Purpose

To develop a practical, easily implementable risk stratification model based on preoperative contrast-enhanced CT (CECT) nodal features to predict the probability of pathologic extranodal extension (pENE) in patients with oropharyngeal squamous cell carcinoma (OPSCC).

Materials and Methods

Preoperative CECT studies in consecutive patients with OPSCC who underwent surgical resection between October 2012 and October 2020 were examined by four neuroradiologists, blinded to the pathologic outcome, for imaging features of pENE. The pathology report was queried for the presence of pENE. Decision tree analysis with cost-complexity pruning was performed to identify a clinically pragmatic model to predict pENE.

Results

A total of 162 patients (median age, 60 years [IQR, 54–67 years]; 134 male, 28 female) with 208 dissected heminecks were included. The primary OPSCC site for most patients was tonsil (67%, 109 of 162) or base of tongue (31%, 50 of 162). Most patients had early-stage disease (American Joint Committee on Cancer Staging Manual eighth edition category T0–T2, 93% [151 of 162]; N0–N1, 90% [145 of 162]). Pathologically confirmed pENE was reported in 28% (45 of 162) of patients. CECT features that were significantly associated with pENE on univariable analysis included size, necrosis, spiculation, perinodal stranding, and infiltration of adjacent structures. Decision tree analysis identified a predictive model including spiculation or irregular margins, matted nodes, and infiltration of adjacent structures. The model had a sensitivity of 41% (19 of 46) and specificity of 96% (157 of 162) for predicting pENE.

Conclusion

The developed model for predicting pENE using preoperative CECT features is practical and had high specificity in patients with OPSCC. Further prospective study is warranted to determine impact on clinical management and outcomes.

Keywords: Head/Neck, CT, Radiation Therapy/Oncology, Neoplasms-Primary, Oncology, Decision Analysis, Observer Performance

Supplemental material is available for this article.

© RSNA, 2025

Summary

In patients with oropharyngeal squamous cell carcinoma undergoing preoperative contrast-enhanced CT, a user-friendly predictive model was developed to identify patients with a high risk of pathologic extranodal extension and to guide oncologic care.

Key Points

• ■ Multiple contrast-enhanced CT features were significantly associated with pathologic extranodal extension (pENE) in patients with oropharyngeal squamous cell carcinoma (OPSCC) treated with surgery, including spiculation (odds ratio, 10.09; P < .001), matted nodes (odds ratio, 7.10; P < .001), and infiltration of adjacent structures (odds ratio, 12.89; P < .001).

• ■ A decision tree model including three nodal CT features (spiculation, matted nodes, and infiltration of adjacent structures) was developed to predict the probability of pENE.

• ■ The proposed decision tree model had high specificity (96%) and may be useful in selecting patients suitable for upfront surgical versus radiotherapeutic management of OPSCC.

Introduction

Oropharyngeal squamous cell carcinoma (OPSCC) is diagnosed in more than 20 000 patients per year in the United States, and the incidence of human papillomavirus (HPV)–associated OPSCC is increasing (1,2). Patients with HPV-associated OPSCC are more likely to experience high rates of disease control and survival with either upfront surgery or radiation therapy (RT). As such, careful selection of upfront surgery or RT is critical to optimize quality of life and functional outcomes (3–6). Patients treated with initial surgery often require risk-adapted adjuvant RT based on pathologic risk factors (eg, advanced primary tumor or nodal classification, perineural invasion, lymphovascular invasion, close surgical margins). Those with positive margins or pathologic extranodal extension (pENE) are treated with chemoradiation therapy (CRT), which is associated with substantial increases in acute toxicity (7–9). For this reason, upfront surgical management is usually not favored when pENE and/or positive margins are expected, unless other patient- or treatment-related concerns (eg, function, wound healing, long-term toxicity) favor surgery. Initial workup of patients with newly diagnosed OPSCC includes comprehensive clinical assessment with physical examination, fiberoptic endoscopy, and diagnostic imaging, including contrast-enhanced CT (CECT) and PET. Reliably identifying pENE during this process is essential, as its presence would be expected to alter the upfront treatment from a surgical to a radiotherapeutic approach to avoid trimodality therapy (ie, surgical resection followed by CRT). Such an approach may avoid the increased rates of acute morbidities associated with CRT, including oral pain, mucositis, dysphagia, odynophagia, feeding tube use, dysgeusia, xerostomia, and dermatitis. The presence of pENE also has implications for deintensification of therapy, which is frequently considered for patients with clinically low-risk OPSCC after surgery. Currently, many deintensification trials exclude patients with pretreatment findings suspicious for clinical or imaging ENE based on clinical examination or diagnostic imaging, respectively (4,10–12).

Multiple imaging features have been associated with pENE after surgical resection (13–15). There is variation in the specific imaging features found to predict pENE across different studies testing the diagnostic accuracy of CECT or MRI for pENE (15). There is also variability in the definition of "positive" findings at preoperative imaging. Many studies define imaging ENE as the presence of a combination of two or more suspicious features (16–18), while others define it as the presence of a single feature (19,20) or do not report the criteria for a positive result (21). Despite the central importance of detecting the presence of pENE prior to treatment selection, no clinically useful predictive systems that could be easily implemented into real-world clinical practice have been developed. While artificial intelligence–based prediction models of pENE based on preoperative CECT continue to mature, these models are challenging to implement in the clinic or may not be readily available, particularly if commercialized (22,23).

To inform multidisciplinary management of patients with OPSCC being considered for upfront surgical versus radiotherapeutic management, we sought to develop a practical, easily implementable risk stratification model based on preoperative CECT nodal features to predict the probability of pENE in patients with OPSCC.

Materials and Methods

Study Sample and Data Collection

This Health Insurance Portability and Accountability Act–compliant study was approved by the institutional review board (approval no. IRB00077854), who granted a waiver of informed consent.

In this retrospective, single-institution cohort study, consecutive patients with OPSCC treated with upfront surgical resection between October 2012 and October 2020 were identified through a systematic query of the electronic medical record for the protocolized College of American Pathologists Cancer Synoptic Reports within surgical pathology reports. A total of 253 patients were identified; patients were excluded for non-OPSCC primary site, recurrent or second primary cancer, no neck dissection performed, or no preoperative CECT performed within 3 months of surgery (Fig 1). Demographics, cancer characteristics, treatment factors, and survival outcomes were extracted from the medical record. Clinical and pathologic staging was performed according to the American Joint Committee on Cancer (AJCC) Staging Manual eighth edition (24). The pathology report, completed as part of routine standard of care, was reviewed from the medical record for results of the nodal dissection, including lymph node involvement, size of the largest involved node, pathologic tumor and nodal classification, pENE, and extent of pENE (in millimeters, where reported).

Figure 1:

Study flow diagram. CECT = contrast-enhanced CT, ND = neck dissection, OPSCC = oropharyngeal squamous cell carcinoma.

Image Acquisition

Considering the retrospective nature of this study, multiple CT scanners were used to acquire the images at our institution or at an outside facility (after which the images were obtained for review by the treating physicians at our institution). As a result, a variety of acquisition parameters were employed. However, most CT images were obtained using 120 kV (peak) and automatic tube current modulation with noise index of 6. The reconstructed section thickness ranged between 0.625 and 3.0 mm (median, 2.0 mm). All images were obtained after the injection of iodinated contrast media, most commonly 90 seconds after injection of 95 mL of iohexol (350 mg of iodine/mL), as lack of preoperative CECT was an exclusion criterion.

Imaging Assessment

Four fellowship-trained neuroradiologists (C.M.L., J.R.S., P.M.B., and K.D.H.) with 9, 7, 6, and 4 years of posttraining experience, respectively, reviewed the preoperative CECT images obtained within the 3 months preceding surgery for each included patient in our institution’s picture archiving and communication system. Neuroradiologists were informed of the laterality of the primary tumor and the side(s) of the neck that underwent neck dissection but were blinded to the surgical pathology report, adjuvant treatment information, and clinical outcomes. For each dissected hemineck, the neuroradiologists independently assessed the presence (yes vs no) of the following imaging features, which were selected as candidate factors for training of the predictive model, based on a literature-informed consensus study team discussion prior to imaging review: central necrosis, perinodal stranding, matted nodes, spiculation or irregular nodal margins, absence of perinodal fat plane, and infiltration of adjacent structures. The neuroradiologists also indicated whether there were multiple nodes (versus a single node) with any of these imaging features present and measured (in centimeters) the axial plane short axis, axial plane long axis, and maximum superoinferior dimension (using either the coronal or sagittal plane, depending on the plane that best depicted the length of the node) of the largest node in each dissected hemineck. As our goal was to develop a practical, straightforward model to predict pENE in a dissected hemineck, neuroradiologists assessed all lymph nodes in each dissected hemineck. Although all neuroradiologists evaluated the same lymph nodes for the presence or absence of the features of interest, the documented assessments were at the level of the hemineck rather than at the level of individual lymph nodes within the hemineck. To maximize interrater reliability during imaging review, previously published exemplary images of the target features were referenced in the materials provided to each neuroradiologist (14). Finally, the neuroradiologists were asked to perform an overall judgment (yes or no) of the presence of the pathologic findings regional nodal metastasis and pENE in each evaluated hemineck.

Statistical Analysis

Descriptive statistics (counts and percentages or medians and 25th–75th percentiles, as appropriate) were calculated for patient and disease characteristics and for the individual and summary imaging assessments. To build a predictive model of pENE, we calculated summary measures of the individual imaging assessments from the four readers. The continuous CECT-based predictors (axial long and short axes, maximum superoinferior dimension, and total number of suspicious features) were each calculated as the average of the four reader assessments. The dichotomous CECT-based predictors (central necrosis, perinodal stranding, matted notes, spiculation or irregular margins, absence of perinodal fat plane, infiltration of adjacent structures, and multiple suspicious features within the hemineck) were each calculated as the majority assessment among the four readers. In the case of ties (two readers noted a presence of the feature and two readers did not), we looked at which of the four readers’ assessments for the presence of pENE were concordant with pathology for that hemineck to approximate the most accurate training data for a predictive model. In these cases, the responses with the highest concordance with the pathologic result were assigned as the summary measure.

Univariable logistic regression models were used to assess the association between the calculated summary CECT-based predictors and presence of pENE in the 208 heminecks. As a sensitivity analysis, univariable mixed-effects logistic regression models with a random intercept for patient and unstructured covariance were performed with the summary CECT-based predictors to control for potential within-patient correlation among the patients with bilateral dissections.

A decision tree was used to predict presence of pENE in the 208 heminecks, using the continuous and dichotomous summary measures of the CECT-based predictors, as well as patient and treatment characteristics. The decision tree model (DTM) was built with entropy criteria, using the gain in information of each factor to determine the splits of the tree, and with the cost-complexity pruning method to prevent overfitting. It was then internally validated using 10-fold cross-validation. Measures of diagnostic accuracy (sensitivity, specificity, misclassification rate, area under the receiver operating characteristic curve, positive predictive value, and negative predictive value) were determined using a threshold probability of pENE of .50. That is, in cases where the DTM predicted the probability of pENE above that threshold, the test was considered positive, and in those where predicted probability of pENE was less than .50, the test was considered negative. The decision tree was internally tested by applying the model to the individual assessments of the four readers. Tests for differences among the sensitivity, specificity, and misclassification rates of the regression tree model applied to the four readers were conducted with four-sample tests for equality of proportions. Fleiss κ was used to measure interrater reliability among the four readers for each CECT-based imaging feature, where the agreement categories include: poor agreement (κ less than 0), slight agreement (0.01–0.20), fair agreement (0.21–0.40), moderate agreement (0.41–0.60), substantial agreement (0.61–0.80), and almost perfect agreement (0.81–1.00) (25). Progression-free survival (PFS) was defined as the time from diagnosis to any disease progression, death, or last follow-up. Overall survival (OS) was defined as time from diagnosis to death from any cause or last follow-up. Rates of adjuvant therapy (none, RT, or CRT) were compared across DTM risk groups using the χ² test. PFS and OS were compared between DTM risk groups using the log-rank test. P values less than .05 were considered statistically significant. Cases with missing data were excluded from each individual analysis. All analyses were performed with SAS version 9.4 (SAS Institute).

Results

Patient Characteristics

In total, 162 patients (134 male, 28 female) with 208 dissected heminecks were included in this analysis. Baseline patient and clinical disease factors are displayed in Table 1. Median age of the cohort was 60 years (IQR, 54–67 years). Most patients identified as White (96.3%, 156 of 162), 63.0% (102 of 162) currently or formerly smoked, and the median pack-years smoking history for those who had ever smoked was 27. Most patients had either tonsil (67.3%, 109 of 162) or base of tongue (30.8%, 50 of 162) primary tumors. HPV status was known (using either p16 immunohistochemistry and/or HPV DNA polymerase chain reaction) in 158 patients and was positive in 142 of 158 (89.9%). The primary tumor clinical stage (according to AJCC eighth edition) was early (T0–T2) in 93.2% (151 of 162) and locally advanced (T3–T4) in 6.8% (11 of 162) of patients. Similarly, clinical nodal classification was limited (N0–N1) in 89.5% (145 of 162) and N2–N3 in 10.5% (17 of 162) of patients. pENE was present in 45 of 162 patients (27.8%) and 46 of 208 (22.1%) heminecks. The median extent of pENE (reported in 10 heminecks) was 1.5 mm (range, 0.6–4.0 mm); three additional neck sides harbored focal pENE, and one had a nodal mass completely replaced by tumor.

Table 1:

Patient and Disease Characteristics

Individual Reader Assessments of Candidate Imaging Factors

Evaluation of 208 heminecks by four independent neuroradiologist reviewers yielded a total of 832 individual hemineck assessments, which are summarized in Table 2. The median axial long- and short-axis measurements were 2.2 and 1.7 cm, respectively. The CECT imaging features of interest were independently assessed by the neuroradiologists from most to least prevalent as: absence of perinodal fat plane (64.8%, 539 of 832), central necrosis (61.5%, 512 of 832), spiculation or irregular margins (27.9%, 232 of 832), perinodal stranding (26.7%, 222 of 832), matted nodes (24.8%, 206 of 832), and infiltration of adjacent structures (8.2%, 68 of 832). Multiple features were present in 40.5% (337 of 832) of assessed heminecks; the median number of suspicious features per assessment was two. Individual readers determined the presence of involved nodal metastases in 76.8% (596 of 776, 56 cases unable to determine) and presence of pENE in 31.5% (261 of 829, three cases unable to determine) of dissected heminecks. For the detection of pathologically positive lymph nodes in the imaged hemineck, the sensitivity and specificity across all four readers was 93% (547 of 590) (95% CI: 90%, 95%) and 74% (137 of 86) (95% CI: 67%, 80%), respectively. The overall sensitivity and specificity for the detection of pENE across all four readers was 60% (111 of 184) (95% CI: 53%, 67%) and 77% (495 of 645, three cases unable to determine) (95% CI: 73%, 80%), respectively. The interrater reliability (measured using Fleiss κ) of each individual CECT feature ranged from fair (0.36 for spiculation or irregular margins) to almost perfect (0.83 for central necrosis), as summarized in Table S1. Table 2 displays the summary measures used to train the predictive model.

Table 2:

Imaging Assessment of CECT Features Selected as Candidate Factors for Predictive Model

Factors Associated with pENE and Diagnostic Accuracy of the DTM

All CECT features assessed were significantly associated with pENE at univariable analysis (Table 3). A sensitivity analysis correcting for within-patient correlations for the patients with bilateral neck dissections found very similar results (Table S2), supporting treating the summary measures for the heminecks as independent observations in a decision tree. Decision tree analysis that considered nodal measurements (axial long axis, axial short axis, and maximum superoinferior dimension) as well as the CECT-based features of interest resulted in a predictive DTM of depth three with six total nodes (Fig 2). The four decision nodes were at spiculation or irregular margins, matted nodes, and infiltration of adjacent structures (yes vs no, in that order). Measures of DTM diagnostic accuracy are presented in Table 4. The DTM accurately classified presence of pENE in 85% of heminecks (176 of 208), with 41% sensitivity (19 of 46), 97% specificity (157 of 162), and an area under the receiver operating characteristic curve of 0.75 (95% CI: 0.68, 0.82). Positive predictive value was 79% (19 of 24), and negative predictive value was 85% (157 of 184). In total, 24 of 208 heminecks were determined to be at high risk (>50%) of harboring pENE by the DTM; 19 heminecks in 18 patients ultimately had pENE at final pathology. For other CECT features (Table 2) not included in the final DTM, the pruning method determined that the gain of information added by the inclusion of each individual feature did not outweigh the resultant increase in DTM complexity. The DTM was internally validated using 10-fold cross-validation: misclassification rate (18%), sensitivity (37%), and specificity (96%) were calculated as averages across the 10 folds.

Table 3:

Univariable Models of Preoperative CECT-based Predictors of Pathologic Extranodal Extension

Figure 2:

Decision tree model predicts pENE by using preoperative CECT features in patients with oropharyngeal squamous cell carcinoma treated with upfront surgery. For each hemineck evaluated, the decision tree model determination of probability of pENE starts with an assessment of spiculation. If absent, the probability of pENE is 12%. If present, an assessment of matted nodes is performed. If matted lymph nodes are present, the probability of pENE is 82%. If matted lymph nodes are not present, the decision tree continues to infiltration. If there is infiltration of adjacent structures, the probability of pENE is 71%. If not, the probability of pENE is 35%. P(ENE) = predicted probability of pENE, pENE = pathologic extranodal extension.

Table 4:

Diagnostic Accuracy of Decision Tree Model for Predicting Pathologic Extranodal Extension Based on Preoperative CECT Imaging

Application of the DTM to Individual Reader Responses

Because the DTM was designed based on summary measures that combined all four radiologists’ responses or, in the case of ties, used the response from radiologists whose estimation of pENE presence (yes or no) was concordant with the pathologic reference standard, we assessed diagnostic accuracy of the DTM when individual radiologists’ responses were applied to the model. The range in calculated specificity was 83%–96% (134 of 161, 143 of 161, 151 of 161, and 156 of 162 for the four readers, respectively) for the DTM (Table 4). Sensitivity was more variable (20%–48%; nine of 46, 13 of 46, 20 of 46, and 22 of 46 for the four readers, respectively), and misclassification rates were higher than in the DTM trained on summary data (20%–26%; 42 of 207, 43 of 208, 43 of 207, and 53 of 207 for the four readers, respectively).

Association between DTM Risk, Adjuvant Therapy, and Clinical Outcomes

We analyzed rates of adjuvant therapy delivery (none, RT alone, combined CRT) by DTM-assigned risk group and found significantly higher rates of CRT delivery in patients determined high risk for pENE (66.7%, 14 of 21) than in patients in the low-risk group (26.8%, 34 of 127; Table S3). There was no evidence of differences between DTM low- and high-risk groups for PFS and OS (Figs S1, S2). PFS at 5 years was 78.6% (95% CI: 71.3%, 86.0%) for low-risk patients versus 75.5% (95% CI: 52.0%, 99.0%) for high-risk patients (Fig S1, log-rank P = .35). OS at 5 years was 78.9% (95% CI: 71.0%, 86.8%) for low-risk patients versus 75.5% (95% CI: 52.0%, 99.0%) for high-risk patients (Fig S2, log-rank P = .15).

Discussion

In this study, we identified multiple CECT-based imaging features associated with pENE in patients undergoing imaging prior to surgical management of OPSCC, including size (long and short axis), necrosis, perinodal stranding, matted nodes, spiculation, absence of perinodal fat plane, and infiltration of adjacent structures. Using these features, we devised a straightforward model including three features (spiculation, matted nodes, and infiltration into adjacent fat planes) to predict the presence of pENE that is also associated with rates of adjuvant CRT to guide risk stratification in the clinic or multidisciplinary tumor board setting. In an era in which patients with clinically limited stage (AJCC eighth edition T1–2 and N0–1) disease have excellent surgical and radiotherapeutic options, patient selection is critical to reduce the incidence of unexpected pENE and minimize the number of patients exposed to the morbidity and quality of life detriments associated with trimodality therapy (6,26). Consistently and accurately identifying patients at high risk of pENE, and subsequently high rates of adjuvant CRT (trimodality therapy), using an easy-to-implement scoring system such as the proposed model would be expected to facilitate more appropriate patient selection for upfront surgical or radiotherapeutic management.

To use this DTM in practice, the reader would assess each individual hemineck, beginning with an assessment of the presence of nodal spiculation. Depending on the presence of each individual feature in the hemineck being assessed, the reader would proceed down the decision tree until reaching a terminal node, at which point, the predicted probability of pENE for that individual hemineck would be displayed. With the high specificity and positive predictive value observed in this model, patients predicted to have a high risk of pENE may be considered for RT over surgery. Patients with a low predicted risk of pENE may be considered good candidates for either initial treatment approach; given the high negative predictive value of the DTM, patients treated with surgery would be expected to have a low rate of pENE. Practically speaking, if this cohort of 162 patients were to have been selected for RT or surgery based on the results of the DTM (using the threshold probability of .50 that was used to define a positive and negative test result), 23 patients (14% of the cohort) with 24 high-risk heminecks (19 of which ultimately were positive for pENE) would have been suggested to be more optimal RT candidates based on the CECT findings.

While several studies have sought to determine the diagnostic accuracy of preoperative imaging to detect pENE, the criteria for diagnosis of imaging ENE and the methods by which such a determination was made vary substantially. In a meta-analysis of the diagnostic performance of CECT and MRI for prediction of pENE in patients with head and neck cancer that included 22 studies, each included study used a different definition of imaging ENE, composed of the presence of one to seven suspicious features (15). While interrater agreement for most of the imaging features found to be predictive of pENE has been shown to be moderate to substantial, consistent with our findings, diagnostic performance would still be expected to vary when using the presence or absence of a single (or multiple) specific feature(s) as the defining criteria for imaging ENE (15,16). This concept is highlighted by the fact that essentially all evaluated imaging features assessed in our study were significantly associated with pENE, but each individual feature had a varying level of influence on the likelihood of pENE. Reliance on the presence of at least one of several factors limits the robustness of the predictive measure; a scoring system encompassing a few of the most statistically influential features may improve diagnostic performance and streamline clinical use. The strength of this study lies in its identification of a straightforward, easily implementable, systematic method to inform providers of the probability of pENE based on preoperative CECT.

Another strength of the present study is its multireader design, which carries implications for the generalizability of our predictive model. Among 22 studies evaluating CECT- or MRI-based prediction of pENE, no studies included more than two readers (15). Our model was trained on summary measures determined from evaluation by four independent neuroradiologists blinded to the pathologic outcome. Even when the model was applied to the individual responses from each radiologist, specificity remained high (>0.83) across all four readers. These findings warrant further validation but strongly suggest generalizability of the identified model to perform similarly across other readers. Additionally, the rate of pENE observed in our study is consistent with prior reports of patients with unselected head and neck squamous cell carcinoma, albeit slightly lower than what may be expected in a population of patients with HPV-associated OPSCC (4,6,17). The prevalence of pENE in this cohort suggests that the patient sample (one of the largest single cohorts reported in the literature on this topic) is representative of the broader head and neck cancer population. Further validation in nonoropharyngeal cancer cases is warranted.

An important theme across the head and neck oncologic and neuroradiologic literature is that patient selection, although critically important, is often complicated by nuance and heterogeneity in management that is informed by clinical examination, baseline functional status, treatment-related functional impacts, preoperative imaging, and disease characteristics (13,27). There are multiple factors to consider when determining the optimal course of treatment for patients with OPSCC, such as oncologic outcome, which is particularly relevant for non–HPV-associated OPSCC, and functional swallowing outcomes (28,29). A tangible benefit of the predictive model developed in this study is the clinically meaningful predicted probabilities associated with each terminal node of the decision tree. While the pENE risk threshold above which a radiotherapeutic approach would be favored over initial surgical management is not firmly established, the model identified in this study pragmatically categorizes patients as low risk (predicted probability of pENE, 12%), intermediate risk (35%), and high risk of pENE (71%–82%). When considering the decision point to proceed with surgical management (or not) based on the CECT-based pENE risk, high specificity (as seen in this model) is a favorable metric from the standpoint of minimizing patient exposure to trimodality therapy in those with a high predicted probability of pENE. When categorizing patients based on low (<50%) and high (≥50%) risk, a significantly greater proportion of patients underwent adjuvant CRT in the high-risk group. Though the proposed DTM was not associated with survival outcomes in our cohort, the observed lack of a significant survival difference more likely represents the fact that patients with pENE underwent additional therapy (ie, adjuvant CRT) that those without pENE did not require. Based on the results of prior trials indicating a benefit to adjuvant CRT and no clear difference in outcomes between either modality, we would expect these patients to have similar survival if they had undergone upfront CRT, while eliminating the morbidity associated with surgery (7–9,26,30).

This study was limited by its retrospective nature, sample size, and rate of observed pENE, particularly pertaining to the number of patients within the individual subsets at the terminal nodes of the decision tree. The study cohort consisting mostly of patients with HPV-associated OPSCC who were deemed good surgical candidates (and treated with upfront surgical resection) may reduce the generalizability of the developed model to patients with non–HPV-associated disease, nonoropharyngeal primary sites, or those who are considered poor surgical candidates. Additional studies are needed to validate this model across a broader range of patients with a variety of disease states. Due to the low number of HPV-negative patients, additional testing within this subgroup was not feasible, but HPV positivity was observed to be a low-importance factor during DTM development. Because extent of pENE was not consistently reported in pathology reports during the study period, only 14 heminecks had pENE extent reported, and no meaningful conclusions were able to be drawn based on this subset. The model developed in this study was based on responses from subspecialty-trained neuroradiologists at a large, tertiary academic center. While we plan to validate the proposed DTM in additional retrospective and prospective observational studies, further study is warranted to confirm similar diagnostic accuracy in the community setting. Similarly, it is unclear whether these findings are applicable to patients with non–squamous cell carcinoma head and neck cancers, and multisite studies may be needed to adequately evaluate its use in these less common head and neck malignancies. Subsequent prospective observational studies, possibly embedded within a routine multidisciplinary head and neck oncology conference, would facilitate validation of the model’s use in the clinic and would also allow the inclusion of patients with other types of primary head and neck cancer.

In conclusion, we developed a pragmatic DTM to predict the probability of pENE in patients with OPSCC, based on CECT features from pretreatment imaging. This simple model that could be quickly administered in routine clinical practice and multidisciplinary tumor boards warrants further validation. Future studies may focus on external validation in OPSCC and non-OPSCC cohorts, ideally through multicenter, prospective investigation.