Magnetic Resonance Imaging‐Based Radiomics Features Associated with Depth of Invasion Predicted Lymph Node Metastasis and Prognosis in Tongue Cancer

Abstract

Background

Adequate safe margin in tongue cancer radical surgery is one of the most important prognostic factors. However, the role of peritumoral tissues in predicting lymph node metastasis (LNM) and prognosis using radiomics analysis remains unclear.

Purpose

To investigate whether magnetic resonance imaging (MRI)‐based radiomics analysis with peritumoral extensions contributes toward the prediction of LNM and prognosis in tongue cancer.

Study type

Retrospective.

Population

Two hundred and thirty‐six patients (38.56% female) with tongue cancer (training set, N = 157; testing set, N = 79; 37.58% and 40.51% female for each).

Field Strength/Sequence

1.5 T; T2‐weighted turbo spin‐echo images.

Assessment

Radiomics models (R_prim, R_prim+3, R_prim+5, R_prim+10, R_prim+15) were developed with features extracted from the primary tumor without or with peritumoral extensions (3, 5, 10, and 15 mm, respectively). Clinicopathological characteristics selected from univariate analysis, including MRI‐reported LN status, radiological extrinsic lingual muscle invasion, and pathological depth of invasion (DOI) were further incorporated into radiomics models to develop combined radiomics models (CR_prim, CR_prim+3, CR_prim+5, CR_prim+10, CR_prim+15). Finally, the model performance was validated in the testing set. DOI was measured from the adjacent normal mucosa to the deepest point of tumor invasion.

Statistical Tests

Chi‐square test, regression analysis, receiver operating characteristic curve (ROC) analysis, decision analysis, spearman correlation analysis. The Delong test was used to compare area under the ROC (AUC). P < 0.05 was considered statistically significant.

Results

Of all the models, the CR_prim+10 reached the highest AUC of 0.995 in the training set and 0.872 in the testing set. Radiomics features were significantly correlated with pathological DOI (correlation coefficients, −0.157 to −0.336). The CR_prim+10 was an independent indicator for poor disease‐free survival (hazard ratio, 5.250) and overall survival (hazard ratio, 17.464) in the testing set.

Data Conclusion

Radiomics analysis with a 10‐mm peritumoral extension had excellent power to predict LNM and prognosis in tongue cancer.

Oral cancer (OC) is a highly heterogeneous disease with aggressive behavior, and accounting for nearly 377,713 new cases and 177,757 deaths worldwide in 2020. ¹ Despite great development in treatment strategies, the 5‐year survival rate of OC patients remains at 50% to 60%. ² , ³ Regional lymph node metastasis (LNM) is considered to be the most important factor contributing to poor prognosis. ³ , ⁴ With the presence of cervical LNM, the survival rate dramatically decreases by approximately 50%. ⁵ Hence, accurate preoperative identification of cervical LN involvement is crucial for optimal treatment strategy and prognosis prediction.

Depth of invasion (DOI) is a well‐known predictor of neck nodal metastasis, local recurrence, and survival in patients with OC. ⁶ , ⁷ , ⁸ Tongue cancer, the most common OC, has a clear anatomical boundary. Because of the specialized function of the tongue (eg, tongue movement due to speech and swallowing), tumor cells tend to migrate along the muscle fibers, making the measurement of DOI at multiple spatial dimensions very important. ⁹ The peritumoral tissues, a crucial component of the microenvironment, carry tumor aggressiveness‐associated information and play an important role in metastasis and recurrence. ¹⁰ , ¹¹ However, it is still a challenging task to effectively find out the “true” margin of the primary tumor and excavate subtle change of peritumoral tissues using conventional medical images, such as magnetic resonance imaging (MRI), computed tomography, and so on. Therefore, a comprehensive evaluation of tumoral and peritumoral features may provide valuable information in the prediction of LNM.

MRI has been routinely used to preoperatively characterize primary tumors and assess clinical staging to establish treatment strategy for patients with OC. Tumor‐associated qualitative variables measured on MRI may contribute to the diagnosis of LN status; however, the reported sensitivity and specificity have been variable. ¹² , ¹³ Moreover, conventional MRI such as T1‐weighted images and T2‐weighted images were insufficient to detect micrometastases of less than 3 mm. ¹⁴ Hence, there is an urgent need to develop a more effective and noninvasive technique for mining information from MRI and facilitating the prediction of LNM.

“Radiomics” refers to a process that extracts and mines quantitative features from medical images via high‐throughput computational techniques, thus uncovering disease characteristics that fail to be detected by the naked eye. ¹⁵ Compared with invasive biopsy, radiomics characterizes complete tumor and its microenvironment heterogeneity in a noninvasive and cost‐effective way without spatial and temporal limitations. Previous radiomics studies in intratumoral and peritumoral regions have shown to predict staging, treatment response, LNM, and survival. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ Furthermore, the radiomics nomogram incorporating radiomics signature and clinicopathological characteristics, including LN stage and molecular subtype, performed better than radiomics signature in disease‐free survival estimation. ²⁰

The aims of this study were: 1) to construct model based on radiomics features extracted from primary tumor to predict LNM of patients with tongue cancer; 2) to explore the incremental value of peritumoral extensions and clinicopathological characteristics in LNM prediction; 3) to assess the predictive power of radiomics model in MRI‐reported LN negative subgroup; 4) to explore the correlation between three‐dimensional (3D) radiomics features and pathological DOI; 5) to investigate the association between the optimal radiomics model and survival outcomes.

Materials and Methods

Patients

This study was approved by our Institutional Review Board with written informed consent waived (ERC‐[2016]‐30). In total, 861 patients who had undergone radical resection glossectomy with or without neck dissection between September 2012 and November 2019 were recruited consecutively. The patient recruitment scheme is presented in Fig. S1 in the Supplemental Material. The inclusion criteria were as follows: 1) patients who had undergone standard head and neck MRI less than 30 days before radical resection; 2) availability of complete clinical and pathological information; and 3) the presence of newly diagnosed and surgically treated tongue cancer. The exclusion criteria were as follows: 1) missing survival information; 2) the presence of a concurrent malignant tumor other than on the tongue; and 3) MRI either degraded or with obvious artifacts. Finally, 236 patients conformed to the inclusion and exclusion criteria, and were randomly divided into two groups, a training set (N = 157) and an independent testing set (N = 79). The two sets were used to train and validate the radiomics models to preoperatively identify patients with high‐risk LNM.

The baseline demographic information was obtained from medical records. To evaluate DOI and tumor budding, hematoxylin and eosin‐stained tumor specimens were reviewed independently by two oral surgeons (J.S.H. and X.Q.L., with more than 20 years of experience) who were blinded to the clinicopathological data. Any disagreement was resolved by consensus. The LN status was determined according to pathological report in patients had undergone neck dissection at initial surgery. For patients had undergone radical excision of the primary tumor only, the LN status was evaluated during 2‐year follow‐up period. Patients who developed suspicious LNM during follow‐up and were subsequently subjected to fine‐needle aspiration, biopsy, or neck dissection. MRI parameters, including DOI, LN status, neurovascular invasion (NVI), extrinsic lingual muscle invasion (EMI), and lingual‐septum invasion (LSI), were assessed blinded by three observers (J.S.H., X.Q.L., and Xiangbo Wan, two oral surgeons and one radiologist with more than 20 years of experience reading MRI). The assessment criteria are described in the Supplementary Information.

MRI Acquisition

Before undergoing radical resection, all patients had undergone standard head and neck MRI with a 1.5T Avanto MR scanner (Siemens Medical Solutions, Erlangen, Germany). Axial T2‐weighted turbo spin‐echo images were retrieved for feature extraction. The image acquisition parameters were as follows: repetition/echo time, 5800/87 or 6500/78 msec; number of averages, 2; echo train length, 13; flip angle, 15 degree; acquisition matrix, 224 * 320; reconstruction matrix, 320 * 320; number of slices, 36; slice thickness, 4.5 or 5 mm; spacing between slices, 6.5 mm; and field of view, 246 * 246 mm².

Region of Interest Segmentation and Image Preprocessing

As depicted in Fig. 1, the contour of the primary tumor and tongue body were manually delineated on consecutive axial slices with the open‐source software ITK‐SNAP (version 3.6.0, www.itksnap.org) by two trained oral surgeons (F.W. and K.F., each with 5 years of experience) independently in a blinded fashion. For avoiding possible dilatation beyond the structure of tongue and mouth floor, we limited the surface of tongue body as the upper boundary. As for the mouth floor, the inner side of musculus mylohyoideus was the lateral boundary, with musculus genioglossus and sublingual gland, tongue base, and hyoid bone as the anterior, posterior, and inferior boundaries, respectively. Then, the ROIs were 3D dilated (3, 5, 10, and 15 mm, respectively) outside the primary tumor with AIMED software (version: 1.6.4, Blot Info & Tech Co. Ltd, Beijing, China). Finally, the extended regions intersected with extension boundaries to produce the true ROIs for the subsequent radiomics analysis. Differences in image acquisition parameters, such as slice thickness, may introduce noise interference into the data. Therefore, the images were normalized in the whole slice images and resampled to a voxel resolution of 1 × 1 × 1 mm³.

Figure 1.

Workflow of region of interest (ROI) segmentation and feature extraction. First, ROIs of the primary tumor and extension boundaries were segmented on T2‐weighted images. Then, the ROIs were three‐dimensional dilated (3, 5, 10, and 15 mm, respectively) outside the primary tumors. Finally, the extended regions intersected with extension boundaries to produce the true ROIs for following radiomics analysis. Three‐dimensional radiomics features, including 540 histogram of oriented gradient (HOG) features, 156 statistical features, 42 texture features, and 48 wavelet features were extracted to quantify shape, intensity, texture, and wavelet.

Feature Extraction and Selection

To extract comprehensive information from the tumor and peritumoral microenvironment, we regarded the ROIs as a whole to extract radiomics features instead of analysis of intratumoral and peritumoral radiomics features separately. Radiomics features were extracted from ROIs with Python (version 3.5.2; https://www.python.org). In total, 786 radiomics features were extracted from each ROI, including 540 histograms of oriented gradient (HOG) features, 42 texture features, 156 statistical features, and 48 wavelet features. The normalization procedure was performed to change the values of extracted features to a range of 0 to 1. To reduce dimensionality and redundant features, a three‐step procedure was performed to select LNM relevant radiomics features from the training set: 1) invalid features, including infinite values, null values, and values with variance of zero were removed; 2) intraclass correlation coefficient (ICC1) and interclass correlation coefficient (ICC2) of each radiomics feature were calculated to quantity feature extraction reproducibility (Supplementary Information). Features with ICC1 and ICC2 values greater than 0.8 were considered reproducible and robust, and therefore retained for analysis; 3) Subsequently, principal component analysis was conducted to select a small number of unique features, which is one of the most widely used data dimension reduction algorithms. The process of principal component analysis was described in the Supplementary Information.

Model Development and Validation

The support vector machine (SVM) was performed on the training set to evaluate and optimize model hyperparameters employing 5‐fold cross‐validation. The radiomics models (R) were developed based on selected radiomics features. The radiomics models with 3D features extracted from the primary tumor or dilated ROIs were called R_prim, R_prim+3, R_prim+5, R_prim+10, R_prim+15, respectively. To evaluate the incremental value of clinicopathological characteristics including age, sex, T‐staging, MRI‐reported LN status, and so on in LNM risk prediction, we conducted a univariate analysis. Characteristics with P values <0.05 were incorporated into R models to build combined radiomics models (CR), which were called CR_prim, CR_prim+3, CR_prim+5, CR_prim+10, CR_prim+15, respectively. Besides, clinicopathological model (CP) was constructed using the selected clinicopathological characteristics. The predictive performance was then validated in the independent testing set.

Statistical Analysis

Statistical analyses were implemented with R software (version 3.5.2; R Foundation for Statistical Computing, Vienna, Austria) and SPSS (version 24.0; IBM Corporation, Armonk, NY, USA). The chi‐square test was performed for the analysis of categorical variables. Patients with a prediction probability greater than 0.5 were considered to be at high‐risk for LNM. The predictive ability of the prognostic models was evaluated by the area under the receiver operating characteristic curve (AUC) values and AUC's 95% confidence interval (CI). The DeLong test was performed to compare AUC values between prognostic models. ²¹ Calibration curves were plotted to compare the predicted and actual values. In addition, accuracy rate, sensitivity, and specificity value were calculated. To calculate the clinical usefulness of the optimal prognostic model, decision curve analysis was conducted to quantify the net benefits at different threshold probabilities in the testing set. ²² Spearman correlation coefficient was used to calculate correlations between pathological DOI and radiomics features. Kaplan–Meier analysis and the log‐rank test were performed to depict and compare disease‐free survival (DFS) and overall survival (OS) curves, respectively. Univariate and multivariate Cox regression analysis was performed to investigate the association between model‐predicted LNM and survival outcomes. A two‐sided P value <0.05 was considered statistically significant.

Results

Patient Characteristics

The clinicopathological characteristics of patients with tongue cancer in the training and independent testing sets are presented in Table 1. Patients of 61.44% (145/236) were men, and the average age of all patients were 50.83 ± 13.58 years (range, 19–86 years). All patients received radical excision of the primary tumors, and 166 underwent neck dissection simultaneously. Of the 99 patients with LNM, 56 patients were diagnosed at surgery. The remaining 43 patients developed LNM during 2‐year follow‐up period. The LN‐positive patients constituted 42.04% (66/157) and 41.78% (33/79) in the training and testing sets, respectively (P = 0.969). Clinicopathological characteristics showed no differences between the two sets, indicating balanced distribution.

TABLE 1.

Baseline Characteristics of Patients in the Training and Testing Sets

Characteristics	Training Set (N = 157)		P Value	Testing Set (N = 79)		P Value*	P Value†
	LN (−) (N = 91)	LN (+) (N = 66)		LN (−) (N = 46)	LN (+) (N = 33)		0.969
Age, years			0.811			0.152	0.978
<60	66	49		31	27
≥60	25	17		15	6
Sex			0.947			0.525	0.663
Female	34	25		20	12
Male	57	41		26	21
T‐staging			0.220			0.058	0.107
T1–T2	49	29		32	16
T3–T4	42	37		14	17
MRI‐reported LN status			0.001			0.001	0.157
Negative	75	39		43	21
Positive	16	27		3	12
Radiological NVI			0.155			0.058	0.446
Negative	70	44		39	22
Positive	21	22		7	11
Radiological DOI			0.240			0.058	0.154
≤10 mm	50	30		32	16
>10 mm	41	36		14	17
Radiological EMI			0.014			0.012	0.207
Negative	76	44		43	23
Positive	15	22		3	10
Radiological LSI			0.995			0.595	0.874
Negative	80	58		42	28
Positive	11	8		4	5
Tumor budding			0.050			0.013	0.088
Low	46	23		31	13
High	45	43		15	20
Pathological DOI			0.006			0.102	0.394
≤10 mm	72	39		38	22
>10 mm	19	27		8	11

LN = lymph node; NVI = neurovascular invasion; DOI = depth of invasion; EMI = extrinsic lingual muscle invasion; LSI = lingual‐septum invasion.

*P value represents the difference in clinicopathological characteristics between LN‐positive and LN‐negative cohorts in the training and testing sets; †P value represents the comparison between training and testing sets.

Tumor Boundary Visualization

As shown in Fig. 2a, the tumor was infiltrated towards the lingual‐septum and mouth floor, which reflects the invasiveness of tongue cancer. The 3D extension toward mouth floor was visualized using anatomical diagram in Fig. 2b. Not only peritumoral tissue on the MR cross section images, but also tissue in mouth floor was included to extract radiomics features. The image of hematoxylin and eosin‐stained tongue squamous cell carcinoma was shown in Fig. 2c, which reflects the infiltrative growth pattern of tumor.

Figure 2.

MR image, anatomical diagram, and H&E‐stained slide of tongue squamous cell carcinoma. (a) The tumor is infiltrated toward the lingual‐septum and mouth floor. (b) The primary tumor and peritumoral extensions are visualized on coronal section using anatomical diagram. (c) The image of H&E‐stained tongue squamous cell carcinoma.

Clinicopathological Model

As shown in Table 1, univariate analysis in the training set showed MRI‐reported LN status, radiological EMI, and pathological DOI were significantly associated with LNM. Multivariate logistic regression analysis identified MRI‐reported LN status (OR 2.432, 95% CI, 1.093–5.411) as an independent predictor of LNM. As shown in Table 2, the MRI‐reported LN status had an AUC of 0.643 (95% CI, 0.545–0.741, Fig. 3a) in the training set, with the accuracy, sensitivity, and specificity of 64.97%, 40.91%, and 82.42%, respectively. In the testing set, the AUC, accuracy, sensitivity, and specificity were 0.649 (95% CI, 0.522–0.777, Fig. 3b), 69.62%, 36.36%, and 93.48%, respectively. The SVM‐based CP model, which included MRI‐reported LN status, radiological EMI, and pathological DOI, yielded an AUC of 0.933 (95% CI, 0.873–0.993, Fig. 3a) in the training set, with the accuracy, sensitivity, and specificity of 86.06%, 84.84%, and 86.87%, respectively. In the testing set, the performance declined, with the AUC, accuracy, sensitivity, and specificity of 0.700 (95% CI, 0.660–0.740, Fig. 3b), 72.15%, 60.61%, and 80.44%, respectively. Nevertheless, the SVM‐based CP model performed significantly better than the MRI‐reported LN status, with the AUC elevated 0.051 (0.700 vs. 0.649) in the testing set.

TABLE 2.

Performances of Prognostic Models in the Training and Testing Sets

Model	Set	AUC	Accuracy (%)	Sensitivity (%)	Specificity (%)
MR	Training	0.643 (0.545–0.741)	64.97 (55.36–73.26)	40.91 (32.68–48.94)	82.42 (73.56–90.02)
	Testing	0.649 (0.522–.0777)	69.62 (60.16–76.38)	36.36 (30.74–44.37)	93.48 (96.33–89.79)
CP	Training	0.933 (0.873–0.993)	86.06 (80.86–91.26)	84.84 (80.84–88.84)	86.87 (88.85–92.89)
	Testing	0.700 (0.660–0.740)	72.15 (66.15–78.15)	60.61 (52.61–68.61)	80.44 (75.43–85.45)
Rprim	Training	0.978 (0.963–0.993)	92.99 (89.15–86.09)	89.39 (85.19–93.59)	95.6 (93.12–98.18)
	Testing	0.710 (0.680–0.740)	73.42 (68.3–78.54)	60.6 (56.00–65.20)	82.6 (78.59–86.61)
Rprim+3	Training	0.978 (0.966–0.990)	94.9 (92.1–97.71)	92.42 (88.42–96.42)	96.70 (94.97–98.46)
	Testing	0.793 (0.765–0.821)	77.21 (73.06–81.36)	66.67 (62.47–70.87)	84.78 (81.27–88.29)
Rprim+5	Training	0.985 (0.975–0.995)	96.17 (93.57–98.77)	93.93 (90.23–97.63)	97.8 (96.32–99.31)
	Testing	0.784 (0.759–0.809)	81.01 (77.01–85.01)	72.72 (68.92–76.52)	86.96 (83.73–90.19)
Rprim+10	Training	0.992 (0.984–1.00)	97.45 (94.95–99.95)	95.45 (92.25–98.65)	98.9 (98.00–99.83)
	Testing	0.855 (0.833–0.877)	86.08 (82.50–89.66)	78.79 (75.59–81.99)	91.30 (88.73–93.87)
Rprim+15	Training	0.987 (0.978–0.996)	95.51 (93.31–97.71)	95.38 (92.38–98.38)	95.60 (93.44–97.83)
	Testing	0.823 (0.795–0.851)	83.54 (79.81–87.27)	75.75 (72.15–79.35)	89.13 (85.88–92.38)
CRprim	Training	0.987 (0.974–0.992)	93.63 (90.73–96.53)	90.9 (86.9–94.9)	95.6 (93.63–97.57)
	Testing	0.720 (0.680–0.760)	78.48 (73.47–83.49)	66.67 (62.57–70.77)	86.96 (83.45–90.47)
CRprim+3	Training	0.988 (0.979–0.997)	96.15 (93.35–98.95)	93.93 (90.43–97.28)	97.77 (96.57–98.96)
	Testing	0.787 (0.757–0.817)	81.01 (76.96–85.06)	69.69 (65.69–73.26)	89.13 (85.93–92.33)
CRprim+5	Training	0.992 (0.985–0.999)	96.81 (94.31–99.31)	95.45 (92.25–98.65)	97.8 (96.75–98.78)
	Testing	0.793 (0.764–0.823)	84.81 (80.96–88.66)	75.76 (72.56–78.88)	91.30 (88.77–93.83)
CRprim+10	Training	0.995 (0.991–0.999)	97.45 (95.35–99.95)	95.45 (93.39–98.76)	98.9 (98.12–99.82)
	Testing	0.872 (0.847–0.897)	87.34 (84.32–90.36)	78.78 (75.78–81.63)	93.47 (91.21–95.73)
CRprim+15	Training	0.991 (0.984–0.998)	95.54 (93.14–97.94)	92.42 (89.92–94.82)	97.80 (96.61–99.01)
	Testing	0.808 (0.788–0.828)	80.24 (76.72–83.76)	68.57 (62.38–74.57)	89.13 (86.32–91.94)

MR = MRI‐reported lymph node status; SVM‐CP = support vector machine‐based clinicopathological model; R = radiomics models; CR = combined radiomics models; AUC = area under the receiver operating characteristic curve.

Figure 3.

Receiver operating characteristic curves of different diagnostic models in the training set (a) and testing set (b). The combined radiomics model of the primary tumor with a 10‐mm peritumoral extension (CR_prim+10) reached the highest AUC of 0.995 (95% CI, 0.991–0.999) in the training set and 0.872 (95% CI, 0.847–0.897) in the testing set. Calibration curves of prognostic models in the training set (c) and testing set (d). The CR_prim+10 showed good agreement between model‐predicted probability and the actual LNM rate. The lines of diagnostic model have a closer fit to the 45° dotted grey line, indicating their better predictive accuracy.

Radiomics Model with Peritumoral Extensions

After feature reduction, top 15, 17, 18, 25, and 10 radiomics features from the primary tumor, and the primary tumor with 3D peritumoral extensions (3, 5, 10, and 15 mm, respectively) were selected to build R models, respectively. The importance of selected features to R models are presented in Figs. S2–S6 in the Supplemental Material. In the training set, the R_prim showed a favorable predictive performance, with an AUC of 0.978 (95% CI, 0.963–0.993, Fig. 3a). After different peritumoral extensions, the R_prim+3, R_prim+5, R_prim+10, and R_prim+15, achieved 0.978 (95% CI, 0.966–0.990), 0.985 (95% CI, 0.975–0.995), 0.992 (95% CI, 0.984–1.000), and 0.987 (95% CI, 0.978–0.996), respectively.

In the testing set, the AUC of the R_prim declined to 0.710 (95% CI, 0.680–0.740, Fig. 3b); however, AUC was still significantly higher than that of MRI‐reported LN status and SVM‐based CP. With the increase of the scale of peritumoral extension, the AUC increased and reached the maximum value at 10‐mm peritumoral extension. Therefore, the R_prim+10 yielded the highest AUC of 0.855 (95% CI, 0.833–0.877), which was significantly higher than the AUC of other radiomics models, including R_prim, R_prim+3 (AUC = 0.793, 95% CI, 0.765–0.821), R_prim+5 (AUC = 0.784, 95% CI, 0.759–0.809), R_prim+15 (AUC = 0.823, 95% CI, 0.795–0.851). Meanwhile, the accuracy, sensitivity, and specificity of the R_prim+10 were higher than the above models with values of 86.08%, 78.79%, and 91.30%, respectively.

Radiomics Model with Clinicopathological Characteristics Integrated

MRI‐reported LN status, radiological EMI, and pathological DOI were integrated into R models to construct CR models. The importance of selected features to CR models is presented in Figs. S7–S11 in the Supplemental Material. In the training set, CR_prim, CR_prim+3, CR_prim+5, and CR_prim+15 achieved AUCs of 0.987 (95% CI, 0.974–0.992), 0.988 (95% CI, 0.979–0.997), 0.992 (95% CI, 0.985–0.999), and 0.991 (95% CI, 0.984–0.998), respectively, while the CR_prim+10 achieved the AUC of 0.995 (95% CI, 0.991–0.999, Fig. 3a), significantly higher than other CR models. In the testing set, the CR_prim+10 still achieved the highest predictive power with the AUC of 0.872 (95% CI, 0.847–0.897, Fig. 3b), which was significantly higher than the AUC of CR_prim (AUC = 0.720, 95% CI, 0.680–0.760), CR_prim+3 (AUC = 0.787, 95% CI, 0.757–0.817), CR_prim+5 (AUC = 0.793, 95% CI, 0.764–0.823), and CR_prim+15 (AUC = 0.808, 95% CI, 0.788–0.828). The accuracy, sensitivity, and specificity of the CR_prim+10 were 87.34%, 78.78%, and 93.47%, respectively, which were consistently higher than other models. The detailed results are summarized in Table 2. The calibration curves of prediction models, especially the CR_prim+10, showed good agreement between model‐predicted probability and the actual LNM rate in both the training and independent testing sets (Fig. 3c,d ). Compared with the R_prim+10, R_prim, CP model, treat‐all, and treat‐none scheme, the CR_prim+10 provided the greatest net benefits for predicting LNM for threshold probabilities of 0.415–0.806 (Fig. 4a).

Figure 4.

(a) Decision curve analysis of prognostic models in the testing set. The x‐axis represents the threshold probability. The y‐axis represents the net benefit. The decision curve demonstrated that in threshold probabilities of 0.415–0.806, applying the CR_prim+10 to predict lymph node metastasis (LNM) adds more benefit than other models. (b) The CR_prim+10 presented a favorable predictive performance in the MRI‐reported LN‐negative subgroup, with an AUC of 0.883. (c) The distribution of the CR_prim+10‐predicted LN status in the MRI‐reported LN‐negative subgroup, indicated that the CR_prim+10 had a good discriminate ability.

Discriminative Ability of the CR_{prim +10} in the MRI‐Reported LN‐Negative Subgroup

Of all cases, 60.61% (60/99) with pathologically positive nodes were understaged and 13.87% (19/137) with pathologically negative nodes were overstaged according to MRI reports, which suggested that MRI reports had high specificity but low sensitivity. In the MRI‐reported LN‐negative subgroup, CR_prim+10 presented a favorable predictive performance with an AUC of 0.883 (95% CI, 0.862–0.902, Fig. 4b). The LN status of 87.5% (56/64) were correctly identified, with only three positive and five negative patients misclassified, showing discriminative capability of the CR_prim+10 (Fig. 4c).

Correlation Analysis

In NVI‐negative subgroup, there was a significant negative correlation between pathological DOI and statistic_1_0_‐1‐Homogeneity (r = −0.157, Fig. 5a). In NVI‐positive subgroup, pathological DOI was negatively correlated with statistic_0_1_0‐MaxProbability (r = −0.278, Fig. 5b). In radiological EMI‐positive subgroup, a significant negative correlation between pathological DOI and statistic_0_1_0‐MaxProbability (r = −0.336, Fig. 5c) was found. In tumor budding number ≥5 subgroup, there was a significant positive correlation between pathological DOI and statistic_1_1_‐1‐ClusterShade (r = −0.185, Fig. 5d). Also, there were significant correlations between pathological DOI and some HOG features (Fig. S12 in the Supplemental Material).

Figure 5.

Spearman correlation analysis showed a significant correlation between pathological depth of invasion and three‐dimensional radiomics features in radiological neurovascular invasion (NVI) negative subgroup (a), radiological NVI positive subgroup (b), radiological extrinsic lingual muscle invasion (EMI)‐positive subgroup (c) and tumor budding number ≥5 subgroup (d), respectively.

Representative LN‐negative and LN‐positive cases were showed in Fig. 6a–f. The signal intensity within ROI was relatively higher in the LN‐positive patients as compared with the LN‐negative patients, which indicated distinct textural pattern on MR images between LN‐negative and LN‐positive patients with tongue cancer.

Figure 6.

Axial MR images of tongue cancer in a 33‐year‐old patient without LNM and a 43‐year‐old patient with LNM, respectively (a,d). Segmentation and three‐dimensional peritumoral extensions of the primary tumor (b,e). Heat map of radiomics feature extracted from the primary tumor with a 10‐mm peritumoral extension showed differences between LN‐negative and LN‐positive patients (c,f). Kaplan–Meier survival analysis indicated that the CR_prim+10‐predicted high‐risk of nodal metastasis had a poor disease‐free survival and overall survival in the testing set (g,h).

The Predictive Value of the CR_{prim +10} in Survival Outcomes

The median follow‐up period was 41 months (range: 4–95 months). The DFS rates and OS rates at 2, 3, and 5 years were 66.67%, 55.70%, 44.79%, and 77.78%, 67.13%, 48.96%, respectively. A significant association between the CR_prim+10 and survival outcome was identified, and the LNM high‐risk group had worse DFS and OS compared with the low‐risk group identified by the CR_prim+10 (Fig. 6g,h). Tumor budding and CR_prim+10‐predicted LN status were incorporated into the multivariate Cox analysis of DFS. MRI‐reported LN status, radiological lingual‐septum invasion, and CR_prim+10‐predicted LN status were incorporated into the multivariate Cox analysis of OS. Multivariate Cox regression analysis showed that the CR_prim+10 was an independent preoperative indicator for poor DFS (hazard ratio, 5.250 [95% CI, 2.152–12.808]) and OS (hazard ratio, 17.464 [95% CI, 3.660–83.330]) (Table 3).

TABLE 3.

Univariate and Multivariate Cox Regression Analysis of DFS and OS in the Testing Set

Characteristics	DFS				OS
	Univariate analysis		Multivariate analysis		Univariate analysis		Multivariate analysis
	Hazard ratio	P value	Hazard ratio	P value	Hazard ratio	P value	Hazard ratio	P value
Age (<60/≥60)	0.465 (0.159–1.354)	0.160			0.743 (0.210–2.64)	0.646
Sex (Female/male)	1.557 (0.672–3.610)	0.302			0.905 (0.322–2.55)	0.850
T‐staging (T1–T2/T3–T4)	1.238 (0.562–2.727)	0.596			1.841 (0.667–5.079)	0.238
MRI‐reported LN status (negative/positive)	1.629 (0.680–3.901)	0.273			3.256 (1.151–9.208)	0.026	0.979 (0.245–3.906)	0.976
Radiological NVI (negative/positive)	0.772 (0.290–2.059)	0.606			1.339 (0.426–4.212)	0.618
Radiological DOI (≤10 mm/>10 mm)	1.238 (0.562–2.727)	0.596			1.841 (0.667–5.079)	0.238
Radiological EMI (negative/positive)	1.170 (0.439–3.120)	0.754			1.897 (0.602–5.980)	0.275
Radiological LSI (negative/positive)	1.559 (0.535–4.545)	0.416			3.201 (1.018–10.068)	0.047	4.501 (0.983–20.599)	0.053
Tumor budding (low/high)	3.255 (1.402–7.559)	0.006	2.389 (1.015, 5.624)	0.046	2.760 (0.943–8.079)	0.064
Pathological DOI (≤10 mm/>10 mm)	0.975 (0.389–2.442)	0.957			0.883 (0.249–3.133)	0.847
CRprim+10‐predicted LN status (negative/positive)	6.230 (2.587–15.004)	<0.001	5.250 (2.152–12.808)	<0.001	15.130 (3.393, 67.472)	<0.001	17.464 (3.660–83.330)	<0.001

DFS = disease‐free survival; OS = overall survival; LN = lymph node; NVI = neurovascular invasion; DOI = depth of invasion; EMI = extrinsic lingual muscle invasion; LSI = lingual‐septum invasion; CR_prim+10 = combined radiomics model with a 10‐mm peritumoral extension.

Discussion

In this study, we developed radiomics models with 3D peritumoral extensions to predict LNM and prognosis preoperatively. Our results showed that the CR_prim+10 had the greatest predictive performance, especially in the MRI‐reported LN‐negative subgroup. The 3D radiomics features were significantly correlated with pathological DOI. Importantly, the CR_prim+10 was an independent indicator for poor DFS and OS.

To develop the radiomics models, we investigated tongue cancer, the most common cancer of the oral cavity. The tongue is a unique muscular organ with legible 3D anatomy architecture, making it relatively leisurely to determine the tongue boundary and tumor border with MR imaging. Therefore, the tongue is an appropriate model to study the infiltrative growth pattern of tumor cells and 3D peritumoral extension of primary tumors using radiomics analysis.

Previous radiomics studies in OC have focused on the primary tumor or LN without taking peritumoral regions into account. ²³ , ²⁴ In our study, we focused not only on the primary tumor, but also the peritumoral extensions. Furthermore, instead of simple 2D peritumoral extensions on the MR cross section images, here, we conducted 3D extensions to extract comprehensive information from the cancer nest and peritumoral microenvironment. Our results verified that a 10‐mm peritumoral extension improved the predictive performance of the radiomics model. Peritumoral tissue has morphological and phenotypic differences from non‐tumor‐bearing healthy tissue. ²⁵ , ²⁶ , ²⁷ These substantial phenotypic and genetic changes are apparent up to 10 mm from the tumor margins. Lymphatic vessel and immune cells in peritumoral regions are an accurate predictor of occult LNM and prognosis. ²⁸ , ²⁹ , ³⁰ Also, a large number of studies have shown the power of peritumoral radiomics features for predicting tumor phenotype, recurrence, LNM, and prognosis. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ Similarity, a recent study demonstrated that features from the intratumoral and 10‐mm peritumoral region provided the best performance for pretreatment prediction of response to chemotherapy in hepatocellular carcinoma. ³¹ Our findings agree with prior studies in that the peritumoral extension could potentially improve the ability of the radiomics model to predict LNM. According to the Eighth American Joint Committee on Cancer Guidelines, a “safe” excision of tongue cancer is routinely recommended at least 15–20 mm away from the primary tumor border. ³² Studies have demonstrated that insufficient excision of tumor tissues can result in a poorer outcome. ³³ , ³⁴ Also, our results verified that the en bloc resection with a margin more than 10 mm was relatively safe for patients with tongue cancer.

Numerous studies have suggested that DOI is an independent predictor of neck nodal metastasis, local recurrence, and survival in patients with OC. ⁶ , ⁷ , ⁸ For this reason, now DOI is included in routinely postoperative pathology reports. ³² In our study, DOI was integrated into radiomics analysis. Results demonstrated that clinicopathological characteristics, including pathological DOI, enhanced the predictive power of radiomics model in LNM prediction. Similarly, previous radiomics studies have shown that clinicopathological characteristics have incremental value in LNM and prognosis prediction. ²⁰ , ³⁵ We further conducted the correlation analysis between pathological DOI and radiomics features. Although no significant correlation was observed in all samples, pathological DOI were significantly correlated with homogeneity, maximum probability, and cluster shade in different subgroups, including radiological NVI negative, radiological NVI positive, radiological EMI‐positive and tumor budding number ≥ 5 subgroups. Homogeneity feature and maximum probability feature have been associated with survival and metastasis in glioblastoma patients. ³⁶ Cluster shade feature has been significantly correlated with recurrence‐free survival in breast cancer. ³⁷ These results indicate that radiomics features could evaluate tumor heterogeneity to some extent.

The traditional MRI reports have the disadvantage of low sensitivity and may lead to delayed treatment and a poor prognosis for OC patients with occult LNM. A previous study has demonstrated that radiomics analysis is able to classify LN status in clinically LN‐negative subgroup. ³⁸ Similarly, in our study, the CR_prim+10 showed good predictive properties in the MRI‐reported LN‐negative subgroup, indicating its ability to detect micrometastasis. The CR_prim+10 greatly improved the sensitivity in LNM prediction, suggesting that patients with tongue cancer would greatly benefit from the radiomics models in terms of treatment strategy.

Limitations

First, this was a retrospective study with relatively small sample size. A prospective multicenter validation with larger sample size is needed. Considering that the tumor border could be easily identified based on the most commonly used T2‐weighted images, we only chose the representative T2‐weighted sequence to build radiomics models. However, this approach may lose complementary information on tumor heterogeneity, which is available from other sequences, such as T1‐weighted sequences, contrast‐enhanced sequences. Third, studies have shown the potential of radiomics integrated with genomics. ³⁹ , ⁴⁰ However, the unavailability of genomic information from this retrospective study hindered us to further enhance our model.

Conclusions

We have proposed a radiomics model with a 10‐mm peritumoral extension and clinicopathological characteristics incorporated to accurately predict LNM and prognosis in tongue cancer. This model exhibited powerful discrimination ability and sensitivity in LNM prediction, especially in MRI‐reported LN‐negative patients.

Conflict of interest

The authors declare no potential conflict of interest.

Abbreviations

3D

three dimensional

AUC

area under the receiver operating characteristic curve

CI

confidence interval

CP

clinicopathological model

CR

combined radiomics models

DFS

disease‐free survival

DOI

depth of invasion

EMI

extrinsic lingual muscle invasion

HOG

histogram of oriented gradient

ICC1

intraclass correlation coefficient

ICC2

interclass correlation coefficient

LN

lymph node

LNM

lymph node metastasis

LSI

lingual‐septum invasion

MRI

magnetic resonance imaging

NVI

neurovascular invasion

OC

oral cancer

OS

overall survival

R

radiomics models

ROI

region‐of‐interest

SVM

support vector machine

Supporting information

Appendix S1: Supporting Information