CT-Based Radiomics Nomogram for Prediction of Progression-Free Survival in Locoregionally Advanced Nasopharyngeal Carcinoma

Figures & data

Table 1 Characteristics of Patients with LA-NPC in the Training and Validation Cohorts

Clinical Factors		Total Set	Training Set	Validation Set	p-value
		(N = 311)	(N = 218)	(N = 93)
Age		44.67±10.25	45.39±10.29	42.99±10.01	0.059
Gender	Male	234 (75.2%)	164 (75.2%)	70 (75.3%)	1
	Female	77 (24.8%)	54 (24.8%)	23 (24.7%)
Smoking	No	203 (65.3%)	143 (65.6%)	60 (64.5%)	0.958
	Yes	108 (34.7%)	75 (34.4%)	33 (35.5%)
Family history of NPC	No	283 (91.0%)	196 (89.9%)	87 (93.5%)	0.418
	Yes	28 (9.0%)	22 (10.1%)	6 (6.5%)
T stage	T1	5 (1.6%)	3 (1.4%)	2 (2.2%)	0.701
	T2	56 (18.0%)	40 (18.3%)	16 (17.2%)
	T3	159 (51.1%)	115 (52.8%)	44 (47.3%)
	T4	91 (29.3%)	60 (27.5%)	31 (33.3%)
N stage	N0	4 (1.3%)	1 (0.5%)	3 (3.2%)	0.091
	N1	125 (40.2%)	85 (39.0%)	40 (43.0%)
	N2	153 (49.2%)	114 (52.3%)	39 (41.9%)
	N3	29 (9.3%)	18 (8.3%)	11 (11.8%)
Clinical stage	C3	199 (64.0%)	145 (66.5%)	54 (58.1%)	0.196
	C4	112 (36.0%)	73 (33.5%)	39 (41.9%)
Treatment	CCRT	143 (46.0%)	101 (46.3%)	42 (45.2%)	0.448
	IC+CCRT	80 (25.7%)	52 (23.9%)	28 (30.1%)
	CCRT+AC	88(28.3%)	65(29.8%)	23 (28.0%)
Progression		88(28.3%)	62(28.4%)	26 (28.0%)	0.931
Follow-up time		75.23±23.85	75.56±23.80	74.48±24.06	0.717

Abbreviations: LA-NPC, locoregionally advanced nasopharyngeal carcinoma; CCRT, concurrent chemoradiotherapy; IC, induction chemotherapy; AC, adjuvant chemotherapy.

Table 2 Coefficient Profiles, Category and Filters of the 20 Radiomics Features Selected by LASSO-Logistic

Radiomics Features	Coefficients	Category	Filters
LLL_firstorder_Median	0.680	Firstorder	Wavelet
1_mm_3D_glszm_LargeAreaHighGrayLevelEmphasis	0.534	Textural	Log_sigma
Glrlm_GrayLevelNonUniformity	0.342	Textural	Gradien
HHH_ngtdm_Busyness	0.323	Textural	Wavelet
Glrlm_HighGrayLevelRunEmphasis	0.225	Textural	Original
HLH_glszm_LargeAreaHighGrayLevelEmphasis	0.164	Textural	Wavelet
HLL_gldm_SmallDependenceLowGrayLevelEmphasis	0.148	Textural	Wavelet
LLL_firstorder_TotalEnergy	0.115	Firstorder	Wavelet
Firstorder_Skewness	0.084	Firstorder	Exponential
LHL_ngtdm_Busyness	0.076	Textural	Wavelet
Glrlm_LongRunHighGrayLevelEmphasis	0.065	Textural	Squareroot
LLH_glcm_Autocorrelation	0.033	Textural	Wavelet
Glszm_HighGrayLevelZoneEmphasis	0.033	Textural	Original
Firstorder_TotalEnergy	0.021	Textural	Square
Gradient_ngtdm_Busyness	3.00E-04	Textural	Gradient
Glcm_Autocorrelation	8.46E-05	Textural	Original
HLL_glrlm_LongRunHighGrayLevelEmphasis	−0.007	Textural	Wavelet
Glszm_SizeZoneNonUniformityNormalized	−0.021	Textural	Exponential
HLH_gldm_DependenceNonUniformityNormalized	−0.101	Textural	Wavelet
HLH_firstorder_Mean	−0.200	Firstorder	Wavelet

Table 3 Performance of Radiomics Nomogram, Clinical Nomogram and TNM Staging System

Cohort	Model	AUC	Sensitivity	Specificity	Accuracy
Training	Radiomics signature	0.897	0.903	0.724	0.755
	Radiomics nomogram	0.925	0.871	0.859	0.711
	Clinical nomogram	0.800	0.919	0.583	0.467
	TNM staging system	0.735	0.694	0.705	0.483
Validation	Radiomics signature	0.856	0.846	0.731	0.763
	Radiomics nomogram	0.873	0.962	0.657	0.742
	Clinical nomogram	0.729	0.654	0.672	0.667
	TNM staging system	0.689	1.000	0.269	0.473

Figure 1 Framework of the study (A) Regions of interest (ROI). (B) Three types of features. (C) Feature selection by T-test and Lasso-logistic regression. (D) Establishment and validation of radiomics signature. (E) Establishment and validation of Nomogram.

Figure 2 LASSO-logistic regression analysis for selection of radiomics features and the distribution of radiomics signature (A) The tuning parameters (λ) in Lasso-Logistic regression were selected by 10-fold cross-validation. When log(λ) is equal to −3.87, the AUC reaches its maximum value. (B) LASSO coefficient profiles of 527 radiomics features. 20 features with non-zero coefficients are selected at the optimal λ. Rad-score distribution of (C) patients in the training cohort and (D) patients in the validation cohort. Red bars show scores for patients were progression-free; green bars show scores for patients who experienced tumor progression or death by any cause.

Figure 3 Establishment of Nomograms. (A) Radiomics nomogram, (B) Clinical nomogram and (C)TNM staging model.

Figure 4 Validation of Nomograms. The ROC curves of three models for (A) The training cohort and (B) the validation cohort. The Calibration curves of radiomics nomogram for (C) the training cohort and (D) the validation cohort. The Calibration curves are close to the standard curves, suggesting that the model has high accuracy in both cohorts. Decision curve analysis (DCA) for the radiomics nomogram and clinical nomogram of (E) the training cohort and (F) the validation cohort. The DCA indicates that radiomics nomogram provide more net benefit than the clinical nomogram and TNM staging model with threshold probability in both cohorts.

Figure 5 Kaplan–Meier analysis for (A) All patients, (B) Training cohort and (C) Validation cohort.