Use of Radiomics Combined With Machine Learning Method in the Recurrence Patterns After Intensity-Modulated Radiotherapy for Nasopharyngeal Carcinoma: A Preliminary Study

Abstract

Objective: To analyze the recurrence patterns and reasons in patients with nasopharyngeal carcinoma (NPC) treated with intensity-modulated radiotherapy (IMRT) and to investigate the feasibility of radiomics for analysis of radioresistance. Methods: We analyzed 306 NPC patients treated with IMRT from Jul-2009 to Aug-2016, 20 of whom developed with recurrence. For the NPCs with recurrence, CT, MR, or PET/CT images of recurrent disease were registered with the primary planning CT for dosimetry analysis. The recurrences were defined as in-field, marginal or out-of-field, according to dose-volume histogram (DVH) of the recurrence volume. To explore the predictive power of radiomics for NPCs with in-field recurrences (NPC-IFR), 16 NPCs with non-progression disease (NPC-NPD) were used for comparison. For these NPC-IFRs and NPC-NPDs, 1117 radiomic features were quantified from the tumor region using pre-treatment spectral attenuated inversion-recovery T2-weighted (SPAIR T2W) magnetic resonance imaging (MRI). Intraclass correlation coefficients (ICC) and Pearson correlation coefficient (PCC) was calculated to identify influential feature subset. Kruskal-Wallis test and receiver operating characteristic (ROC) analysis were employed to assess the capability of each feature on NPC-IFR prediction. Principal component analysis (PCA) was performed for feature reduction. Artificial neural network (ANN), k-nearest neighbor (KNN), and support vector machine (SVM) models were trained and validated by using stratified 10-fold cross validation. Results: The median follow up was 26.5 (range 8-65) months. 9/20 (45%) occurred in the primary tumor, 8/20 (40%) occurred in regional lymph nodes, and 3/20 (15%) patients developed a primary and regional failure. Dosimetric and target volume analysis of the recurrence indicated that there were 18 in-field, and 1 marginal as well as 1 out-of-field recurrence. With pre-therapeutic SPAIR T2W MRI images available, 11 NPC-IFRs (11 of 18 NPC-IFRs who had available pre-therapeutic MRI) and 16 NPC-NPDs were subsequently employed for radiomic analysis. Results showed that NPC-IFRs vs. NPC-NPDs could be differentiated by 8 features (AUCs: 0.727-0.835). The classification models showed potential in prediction of NPC-IFR with higher accuracies (ANN: 0.812, KNN: 0.775, SVM: 0.732). Conclusion: In-field and high-dose region relapse were the main recurrence patterns which may be due to the radioresistance. After integration in the clinical workflow, radiomic analysis can be served as imaging biomarkers to facilitate early salvage for NPC patients who are at risk of in-field recurrence.

Keywords: intensity-modulated radiotherapy; nasopharyngeal carcinoma; prediction; radiomic analysis; recurrence pattern.

Figures

Figure 1

Patterns of failure for patients with recurrence, with the accumulated dose and site of recurrence. (A) In field. (B) Out of field. (C) Marginal.

Figure 2

Flowchart of using radiomic analysis in recurrent pattern.

Figure 3

(A) Workflow of radiomic analysis for discrimination between NPC-IFR (NPC with in-field recurrence) and NPC-NPD (NPC with non-progression disease). I, Image segmentation was performed on SPAIR T2W MR images. II, Features were extracted from the tumor contours on the MR images using shape, first order, texture, LoG and wavelet-based method. III, Principal component analysis (PCA) was performed on significant features for dimension reduction. IV, For the analysis, principal components derived from significant features were combined with supervised machine learning method for prediction of NPC-IFR vs. NPC-NPD. (B) Examples of feature maps computed from two-dimensional tumor region by using GLCM method (e.g., Energy, Entropy, Correlation, InverseDifferenceMoment [IDM]).

Figure 4

Box plots of amplitude features, successfully differentiating NPC-IFR from NPC-NPD. (A) glcm_CT (P = 0.046); (B) WHLL_gldm_DE (P = 0.023); (C) WHLH_F_RMS (P = 0.023); (D) WHLL_glcm_CP (P = 0.032); (E) WHLL_ngtdm_Complexity (P = 0.041); (F) WHLH_glcm_IMC (P = 0.041); (G) WHLL_gldm_SDLGLE (P = 0.048); (H) WLLH_ngtdm_Strength (P = 0.048).

Figure 5

Pearson correlation coefficient of the eight significant features.

Figure 6

(A) Receiver operating characteristics (ROC) curves on the basis of the significant features. (B) Three-dimensional scatter plot of the NPC-IFR and NPC-NPD by using three principal components derived from the above eight significant features.