Patient characteristics
The characteristics of patients in both cohorts are summarised in Table 1. Of the 120 patients in the primary cohort, 49.2% (59/120) had cN0 T1 tumors, and 50.8%(61/120) had T2 tumors. The treatment modality was recorded for all patients, 120 patients were treated with primary resection and END, 26.7% (32/120) of whom had nodal metastasis (pN+). The mean values of aTT, cTT, sTT, PLD and SLD was 9.89± 4.07 mm, 13.44± 5.63mm, 14.35± 6.52mm, 4.38 ± 2.47 mm and 3.80 ± 2.94 mm, respectively.The mean follow-up time was 45 months (18–82 months). Recurrence or metastasis occurred in 15% (18/120): at the primary site in 5% (6/120), in the neck in 4.17% (5/120) and 5.83% (7/120) had distant metastasis in the primary cohort.
For postoperative pathologic results,the mean value of pathologic thickness was 8.06 ± 3.47 mm.Pearson’s correlation coefficient showed a strong positive correlation between pathologic thickness and MRI thickness in mediolateral direction (aTT) with 0.836 (P< 0.001). Furthermore, we found a significant difference in predicting cervical lymph node status when incorporating pDOI according to the AJCC (8th edition), no occult metastasis was detected in pDOI ≤ 5mm, metastasis rate was 25.3% in pDOI> 5 mm to ≤ 10 mm and metastasis rate was78.6% in pDOI> 10 mm (P < 0.001). The presence of PNI was also significantly different according to nodal status (P = 0.046).In contrast, cancer cell differentiation (P = 0.824), presence of LVI (P = 0.663) and muscular invasion (P = 0.304) were not significantly different for predicting nodal metastasis.
Identifying cutoff values of MRI parameters to estimatepN+
In this study, we focused on assessment of occult metastasis risk in early-stageSCCLT by evaluation of MRI parameters in three-dimensional planes. ROC curve analysis was applied to calculate the optimal cutoff value of the preoperative MRI parameters for identifying regional lymph node metastasis.The ability and effectiveness of each parameter were analyzed using their cutoff value to calculate sensitivity, specificity, PPV, NPV and likelihood ratio. The effect of cutoff value impacting the diagnostic accuracy is presented in Table 2 and the ROC curve regarding the usedcutoff value of predictors to assess subclinical nodal metastases are also shown in Figure2.
ROC curve analysis indicated that the aTTof 8.5 mm, cTT of 11 mm and sTT of 12 mm might be the optimal threshold value for predicting cervical lymph node metastases, with the AUC being 0.895, 0.867 and 0.916, respectively. Notably, the predictivevalue of aTT with 8.5 mm cutoff value was 84.4% sensitivity, 86.4% specificity, 69.2% PPV, 93.8% NPV, +LR 6.00 and -LR 0.19 (P < 0.001). Likewise the diagnostic value of sTT with 14 mm cutoff value was 90.6% sensitivity, 81.8% specificity, 64.4% PPV, 96% NPV, +LR 5.06 and -LR 0.11(P < 0.001). Additionally,the incidence of nodal metastasis was higher in patients with significantrTTT than in patients with minimalrTTT (65.1% versus 5.2%, P < 0.001). The ROC curves also reflected that PLD of 5 mm and SLD of 4 mm might be the optimal decision threshold related to nodal metastases, with the AUC being 0.759 and 0.616, respectively.
Development of multiparametric MRI-derived nomogram for predicting of individualized occult metastasis
A univariate analysis was initially performed to evaluate the association between each variable and lymph node metastases. Besides,rTTT, PLD and SLD based on MRI, some other risk variables were evaluated, including sex, age, tumor size (T1 or T2), pathologic differentiation (well or moderate-poor), and tumor location (anterior-middle or posterior). The probability of having lymph node metastasis in each variable was estimated. Then, all variables with a P value <0.05 in the univariate analysis were entered in a multivariate logistic regression analysis model by using a backward condition method and remaining variables were used to create a nomogram to predict the probability of cervical lymph node-positive disease in the primary cohort. Multivariate logistic regression analysis demonstratedthat tumor size (OR 15.175, 95% CI 1.436-160.329, P = 0.024), rTTT (OR 11.528, 95% CI 2.483-53.530, P = 0.002), PLD (OR 11.976, 95% CI 1.981-72.413, P = 0.005), and tumor location (OR 6.311, 95% CI 1.514-26.304, P = 0.011) emerged as significant predictors to indicate the existence oflymph node metastasis of early-stage SCCLT(Table 3).
Ultimately, four potential predictors were presented in the final nomogram that identify the occult metastasis risk and predict the patient on the likelihood of cervical nodal involvement (Figure 3). Nomogram was characterized by one scale corresponding with each variable, a score scale, a total score scale, and a probability scale. The top row shows the point assignment for each predictor. The metastasis risk predictors in order of their listing in the nomogram are: tumor size, tumor location, rTTT and PLD which is each predictorrepresent a point value based on the tumor characteristics (row 2-4). The points assigned to each of the four predictors are summed, and the total points indicated in row 5. The bottom row shows the probability of the patient having cervical lymph node metastasis.
Nomogram use for classifying patient metastasis risk group
ROC curve analysis was also applied to identify the optimal cutoff value of total score obtained from the nomogram in the primary cohort.The ROC curve analysis indicated that the optimal cutoff value of nomogram total score of 210 points was significantly different to discriminate the patients likelihood of having cervical lymph node metastasis preoperatively. With regard to this cutoff value, patient may be categorized as having a low occult metastasis risk depending upon their total risk score less than 210 points. While, total risk scores more than 210 points, patients were stratified in a high occult metastasis risk. The possibility of metastasis ratehigher 20% whentotal score more than cutoff value of 210 points was presented.
In the primary cohort, the MRI-based nomogram demonstrateda predictive accuracy of predicting individualized positive cervical lymph nodes, resulting in excellent discrimination power of the AUC 0.952with 95% CI 0.917–0.987, P < 0.001 in Figure 4A. The sensitivity, specificity, PPV, and NPV of the nomogram using the cutoff value of 210 points were 93.8%,87.5%, 73.2% and 97.5%, respectively (P< 0.001) The Hosmer–Lemeshow goodness-of-fit test indicated that the nomogram was well-calibrated and the slope hadno apparent departure from perfect fit(P <0.001). The calibration plot showed good agreement between the prediction and metastasis risk observation (Figure 4C).
Clinical validation of MRI-based nomogram
Validation cohort analysis was carried out on41 patients who were diagnosed as having T1 tumors 51.20% (21/41), and T2 tumors 48.80% (20/41). All patients were treated with primary resection and END, 29.3% (12/41) of whom had cervical node metastasis (pN+).Considering the nomogram cutoff value of 210 points generated the patients into metastasis risk groups, 25 patients (60.98%) with total score obtained from nomogram less than 210 points had a low metastasis risk. On the other hand, 16 patients (39.02%) with total score obtained from nomogram more than 210 points had a highmetastasis risk.Incidence of nodal metastasisdemonstrated a statistically significant difference between the low and high metastasis risk groups (P = 0.001). The incidence of nodal metastasis was higher in the high metastasis risk group (62.5%), than in the low metastasis risk group (8.0%)
Validation of the nomogram was conducted both internally and externally, as presented in Figure 4. In the entire validation cohort, theROC curve analysis reflected the MRI-based nomogram demonstrated excellent prediction performance with high predictive value of AUC 0.881, 95% CI 0.779–0.983 (P<0.001) in Figure 4B. Additionally, we conducted a diagnostic test by using nomogram cutoff value of 210 points, the result showed sensitivity of 83.3%, specificity of 79.3%, PPV of 62.5%, and NPV of 92.0% in the validation cohort. The calibration curves of the nomogram in the validation cohorts also wereplotted, and the slope was close to 1, indicating that the nomograms were well calibrated in Figure 4D.