Abstract
Purpose
Previously, the evaluation of xerostomia depended on subjective grading systems, rather than the accurate saliva amount reduction. Our aim was to quantify acute xerostomia with reduced saliva amount, and apply radiomics, dose-volume histogram (DVH) criteria and clinical features to predict saliva amount reduction by machine learning techniques.
Material and methods
Computed tomography (CT) of parotid glands, DVH, and clinical data of 52 patients were collected to extract radiomics, DVH criteria and clinical features, respectively. Firstly, radiomics, DVH criteria and clinical features were divided into 3 groups for feature selection, in order to alleviate the masking effect of the number of features in different groups. Secondly, the top features in the 3 groups composed integrated features, and features selection was performed again for integrated features. In this study, feature selection was used as a combination of eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) to alleviate multicollinearity. Finally, 6 machine learning techniques were used for predicting saliva amount reduction. Meanwhile, top radiomics features were modeled using the same machine learning techniques for comparison.
Result
17 integrated features (10 radiomics, 4 clinical, 3 DVH criteria) were selected to predict saliva amount reduction, with a mean square error (MSE) of 0.6994 and a R2 score of 0.9815. Top 17 and 10 selected radiomics features predicted saliva amount reduction, with MSE of 0.7376, 0.7519, and R2 score of 0.9805, 0.9801, respectively.
Conclusion
With the same number of features, integrated features (radiomics + DVH criteria + clinical) performed better than radiomics features alone. The important DVH criteria and clinical features mainly included, white blood cells (WBC), parotid_glands_Dmax, Age, parotid_glands_V15, hemoglobin (Hb), BMI and parotid_glands_V45.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Xerostomia is a common side effect of nasopharyngeal carcinoma (NPC) radiotherapy (RT) [1, 2]. It seriously affects the quality of life (QOL) of patients, including oral infection, eating problem, undernutrition, and insomnia [3,4,5]. Intensity modulated radiation therapy (IMRT) is a technique to better protect the organs at risk (OARs) [6], but parotid glands (PGs) are inevitably included in the irradiation field, causing radiation injury. Accurate prediction could assist early intervention of xerostomia.
It is widely reported that radiomics features improve the performance of xerostomia prediction. Radiomics features of PGs extracted from CT [7,8,9,10,11,12], CBCT [13], MRI [14,15,16], and PET [17, 18] reflect the condition of PGs and can be used as an important factor in xerostomia prediction. Van Dijk et al. revealed that PGs surface reduction was associated with late xerostomia [8]. Pota et al. used radiomics features extracted from CT to predict PGs contraction and explored the relationship between PGs contraction and xerostomia [10]. They found that acute xerostomia was positively associated with PGs contraction, while xerostomia after 2 years of RT is negatively associated with it. Wu et al. and Rosen et al. both pointed out that the mean Hounsfield units (HU) of PGs was effective for the prediction of xerostomia [13, 19]. It is known that the mean HU of PGs mirrored the density of PGs. Similarly, Belli et al. revealed that the gradient of PGs density played an important role in predicting xerostomia after RT [20]. Zhang et al. found that the maximum apparent diffusion coefficient of PGs was a sensitive indicator for salivary gland dysfunction, so it was potential to predict xerostomia after RT [21].
Dose-volume histogram (DVH) criteria features were one of the first factors used to predict xerostomia after RT. Many studies showed that the degree of xerostomia after RT decreased as the mean dose of PGs reduced [22,23,24,25,26]. Deasy et al. found that PGs function had the least reduction when the mean dose was less than 10–15 Gy [23]. Miah et al. pointed out that whether the mean dose of PGs is greater than 26 Gy can be used to predict xerostomia after RT [26]. Characteristics extracted from DVH, including V15–V45 and D10-D90, were used to predict xerostomia after RT with a good performance [13, 27, 28]. Gabry et al. defined the change of mean dose along the coordinate axis as the dose gradient, and applied machine learning technologies to build a superior xerostomia prediction model [29]. Through statistical analysis, it was found that some clinical features are significantly different between patients with xerostomia and normal patients, including age, sex, tumor site, chemotherapy [13, 25].
In above studies, the end point of xerostomia was estimated by grading systems which depends on observer or patient self-evaluation, e.g., EORTC QLQC-30 and H&N35 questionnaire, RTOG criteria. EORTC QLQC-30 and H&N35 questionnaire contains dozens of questions, and uses a 4-point Likert scale to describe the condition as ‘none’, ‘a bit’, ‘quite a bit’, and ‘a lot’ [30]. According to the RTOG criteria, the severity of xerostomia is graded to 5 levels from G0 to G4 [31]. However, the results of the above grading system are qualitative and subjective. Besides, it is not straightforward to compare grades from different systems, because the consistency between them may be low [32]. Additionally, it is suggested that EORTC/RTOG is prone to misinterpretation and omission errors, which underestimates the severity of xerostomia [23, 33, 34]. Thus, we believed that it is more objective and accurate to use saliva amount (SA) reduction to quantify xerostomia. In this study, SA reduction was defined as the stimulated SA difference between 0th fraction and 30th fraction. Moreover, acquiring important features predicting SA reduction is conducive to early intervention of xerostomia after RT.
In previous studies, the evaluation of xerostomia depended on subjective grading systems, rather than the accurate SA reduction. In this study, we used radiomics, DVH criteria and clinical features to establish prediction models for SA reduction. To our knowledge, this is the first study to add DVH criteria and clinical features into radiomics model for predicting SA reduction.
2 Materials and method
2.1 Materials
In this study, CT image, DVH, clinical data, and SA0f-30f reduction were collected from 52 NPC patients (24–80 years, median 48 years) who received IMRT in the Sun Yat-sen University Cancer Center (SYSUCC). All patients received radical dosage with a prescription dose of 68.1 Gy in 30 fractions. CT images of all patients at the 0th fraction (planning CT, 0f) and the 10th fraction (10f) during RT were acquired through the CT simulator. PGs contour were delineated on each CT by a radiation oncologist and independently verified by another radiation oncologist. The clinical features included gender, age, tumor stage, BMI, hematological test, blood pressure, etc. Stimulate saliva collection lasted for 5 min at 0th fraction and 30th fraction during RT. Among the 52 patients, 50 patients had reduced SA reduction (maximum SA reduction: 26.2 ml, minimum SA Reduction: 0.1 ml, standard deviation: 6.02), one patient had no change in SA and another patient had increased SA by 0.4 ml. The patients with stages I–IVB, were randomly chosen from the control group in the NPC clinical trial, which was registered on the clinicaltrials.gov (ID: NCT01762514). The study was approved by the institutional review board and the ethical review office from the institution, and the data was submitted to the public scientific research data storage platform (www.researchdata.org.cn). The approval number is RDDB2018000256.
In this study, all CT images were acquired through the CT simulator (Brilliance™ CT, Philips, The Netherlands). The detailed parameters for these protocols were given as following: voltage 120 kVp, exposure 300 mAs, slice thickness 3 mm, increment 3 mm, collimation 16 mm × 0.75 mm, display FOV 600 mm, scan FOV 600 mm, reconstruction filter type UB/B, and pitch 0.567. The DICOM matrix size is 512 × 512 × 113, and the voxel size is 0.9765625 mm × 0.9765625 mm × 3.0000000 mm.
2.2 Feature extraction
The radiomics module of 3D Slicer was used to extract the radiomics features of CT. It had an interface to the PyRadiomics which was an open-source package in python for extracting radiomics features from medical images. With 3D Slicer, we obtained the results of the CT image after the wavelet transform (all combinations of high or low pass filters in each three-dimensional space). Using the radiomics module, we extracted texture features from CT images, including First Order, Shape, Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Neighbouring Gray Tone Difference Matrix (NGTDM) and Gray Level Dependence Matrix. In this study, we used version 3.0 of PyRadiomics. The initial setting parameters of PyRadiomics are as follows: Spatial Resampling = None, Intensity Rescaling = None, Intensity Discretization: binwidth = 25. And the set wavelet type was haar. All other parameters are default. A total of 3404 dimensions of radiomics features were extracted from all results from 0 and 10f CT and CT wavelet transforms. DVH criteria and clinical features are depicted in Table 1. In Table 1, PGs_V n means the volume proportion of the PGs received by n Gy. PGs_Dmax, PGs_Dmean and PGs_Dmin are the maximum, mean and minimum values of the dose of PGs respectively. Take PGs_V45 as an example, PGs_V45 represents the left and right PGs, and the volume receiving 45 Gy dose accounts for the proportion of the total volume of the PGs. PGs_Dmax reflects the maximum dose to the whole PGs.
2.3 Feature selection
In this study, there were high dimensional features. If all features were used for modeling, the generalization of the model will be limited. Thus, feature selection was essential for developing the generalization. Besides, in this study, there were great differences in the order of magnitude of radiomics, DVH criteria and clinical features (radiomics: 3404, DVH criteria: 9, clinical: 15). If the 3 groups of features are selected together, radiomic features will obscure the importance of DVH criteria and clinical features. Therefore, we referred to the practice of Yu et al. [35], and carried out feature selection on these 3 groups of features respectively. In previous studies, the most common feature pre-selection was to retain the features with the highest correlation with the predicted target, in the interval where Pearson correlation was more than 0.8 [7, 15, 17]. Since the Pearson correlation method cannot provide more information between the features and the predicted targets, eXtreme Gradient Boosting (XGBoost) + SHapley Additive exPlanations (SHAP) was chosen to select the features and explore the features influence on the predicted target, even if that feature was not used for modeling finally.
The procedures of XGBoost + SHAP are as follows. First, the prediction models of SA reduction were established by applying XGBoost to radiomics, DVH criteria and clinical features, respectively. Second, for each of the three models, SHAP analysis was used to obtain the weights of feature importance within the groups. Third, with the guidance of radiation oncologists, the top-ranked features of the three groups were selected. The principle of selection was to include most of the feature importance with fewer features. Fourth, the selected radiomics, DVH criteria and clinical features composed integrated features. For integrated features, another model was established by XGBoost, which was analyzed using SHAP then. The purpose was to obtain the feature importance of integrated features and further select the features. The principle of selection was consistent with step 3. The flowchart was shown in the yellow dotted box in Fig. 1.
2.4 XGBoost
Similar to gradient boosting decision tree (GBDT), XGBoost continuously reduces the bias of the additive model by fitting previous base model prediction errors with the new base model. Compared with the traditional GBDT, XGBoost takes the loss function to a second-order Taylor expansion instead of a first-order one, while adding regularization term to the loss function [36]. As a result, XGBoost has better predictive and generalization capabilities. Bianchi et al. found that XGBoost model can diagnose earlier osteoarthritis of the temporomandibular joint, based on biomarkers [37]. The loss function is shown in Eq. (1).
wherein, \({g}_{i}={\partial }_{{\widehat{y}}^{(t-1)}}l({y}_{i},{\widehat{y}}^{(t-1)})\), \({h}_{i}={\partial }_{{\widehat{y}}^{(t-1)}}^{2}l({y}_{i},{\widehat{y}}^{(t-1)})\), are respectively the first and second partial derivatives of the loss function. \(\gamma T+\frac{1}{2}\sum_{j=1}^{T}{{w}_{j}}^{2}\) is the regularization item.
2.5 SHAP analysis
Inspired by cooperative game theory, SHAP [38] analysis constructs an additive explanatory model, which obtain the independent importance of each feature. Considering the influence and synergy between variables, it shows the average contribution margin of each feature, which effectively avoid multicollinearity. SHAP can not only obtain the overall importance of a feature, but also be used to reflect the influence of a feature in different samples. For each prediction sample, SHAP analysis produces a predicted value. It is the sum of the values assigned to each feature, as shown in Eq. (2).
wherein,\({y}_{i}\) is the SHAP value of the model for the \(i\) th sample, and \({y}_{base}\) is the mean SHAP value of all samples. \({x}_{i,j}\) means the \(j\) th feature of the \(i\) th sample.\(f\left({x}_{i,j}\right)\), is the SHAP value of \({x}_{i,j}\), reflecting the contribution of \({x}_{i,j}\) to the final SHAP value \({y}_{i}\). When \(f\left({x}_{i,j}\right)\)>0, it indicates that the feature improves the final SHAP value \({y}_{i}\), i.e., the positive effect; When \(f\left({x}_{i,j}\right)\)<0, it indicates that this feature reduces the final SHAP value \({y}_{i}\), i.e., the negative effect.
2.6 Prediction model
After feature selection, SA reduction model was established by a variety of machine learning technologies to find the most accurate model. Besides, comparing the performance differences of models established by different algorithms and different feature groups, the influence of DVH criteria and clinical features on predicting SA reduction was explored.
2.7 Ridge regression
Ridge adds L2-norm [39] to linear regression loss function, which is beneficial to alleviate the problem of multicollinearity and overfitting. Using radiomics features, Wei et al. used Ridge to predict the prognosis of skull base chordoma [40]. The loss function of ridge is shown in Eq. (3).
wherein, \(\lambda {||\upomega ||}_{2}^{2}\) is the L2-norm. As \(\lambda\) increase, the generalization capability of the model is strengthened.
2.8 Support vector regression (SVR)
SVR seeks a regression hyperplane that minimizes the distance of all the data from that hyperplane [41]. It is suitable for cases with linearly indivisible samples and high characteristic dimensions.
2.9 Decision tree
Decision Tree constructs binary trees to make decisions [42]. It is adept at dealing with the nonlinear relationship and outliers in the data. Sakai et al. used decision tree to detect multi-leaf collimator (MLC) modeling errors with the use of radiomic features [43].
2.10 Random forest
Random Forest is an algorithm integrated by bagging of multiple decision trees [44]. There is no correlation between each decision tree, and the final output of the model is jointly determined by every decision tree. Since the feature subset of each decision tree is randomly selected, it has good robustness and can maintain accuracy even if data are missing. Homayounieh et al. used random forest to differentiate diffuse liver diseases on non-contrast CT [45].
2.11 Adaboost
AdaBoost is a boosting ensemble algorithm with two characteristics [46]. First, samples predicted wrong by previous base model are given high weight to increase the attention of the next base model.
Second, the base models with high accuracy are given higher weight, and the final output is the weighted average output of multiple base models. As a result, AdaBoost has an outstanding prediction performance. Thongkam et al. predicted breast cancer survivability by AdaBoost [47].
3 Result
In this study, the definition of SA reduction is shown in Eq. (4). First, radiomics, DVH criteria and clinical features were modeled by XGBoost, and the three models were analyzed by SHAP. The result was shown in Fig. 2.
In Fig. 2. (c), for PGs_V45, the blue dots are clustered to the left of the longitudinal axis. Blue indicates that the sample has a small value of PGs_V45. The points to the left of the axis all have SHAP values less than zero, indicating that they have the negative effect on the prediction. Therefore, it can be concluded that PGs_V45↓, (SA0f–SA30f) ↓, SA30f ↑, severity of xerostomia ↓. Meanwhile, a lot of red dots are concentrated on the right side of the vertical axis. The red color indicates that the value of PGs_V45 for this sample is large. The dots to the right of the axis, which have SHAP values greater than zero, have a positive effect on the prediction. It is indicating that PGs_V45 ↑, (SA0f–SA30f) ↑, SA30f ↓, severity of xerostomia ↑. Generally, PGs_V45 is positively correlated with the severity of xerostomia. Similarly, in Fig. 2e, it is suggested age was negatively correlated with the severity of xerostomia.
Figure 2b, d, f show the weights of important features in radiomics, DVH criteria, and clinical features, respectively. With the participation of radiation oncologists, the top 10 radiomics, the top 5 DVH criteria, and the top 5 clinical features were selected for the second feature selection. The reason is that fewer features can effectively avoid overfitting. We expect to retain no more than 10 radiomic features to reduce the risk of overfitting. In addition, due to the close number of DVH criteria and clinical features, we set them to equal weight. Meanwhile, in order to investigate the influence of adding DVH criteria and clinical features based on radiomic features, we set the total number of DVH criteria and clinical features consistent with radiomic features. Therefore, from Fig. 2b, d, f, the top 10 radiomics features, the top 5 DVH criteria features and the top 5 clinical features respectively contain most of the importance of the features, which reflects most of the information in all features.
The above 20 features were selected again by XGBoost + SHAP, and the rank of feature importance was shown in Fig. 3. The top 17 features (shown in the yellow dotted box in Fig. 3) were selected while the top 10 radiomics features were included in them. We calculated the Pearson correlation of these 17 features, and the results are shown in Fig. 4.
To avoid multicollinearity, the most commonly method was to retain the feature with the highest correlation to the predicted target among the features with Pearson correlation greater than 0.8. In Fig. 4, it is observed that the Pearson correlations between any features selected using XGBoost + SHAP are less than 0.8. Among the 17 features, the Pearson correlation between “0f_wavelet_lhh_glszm_small-areahighgraylevelemphasis_PR” and “0f_wavelet_hlh_firstorder_minimum_PR” has the largest Pearson correlation of 0.549.
It is concluded that XGBoost + SHAP can effectively avoid feature multicollinearity. Then, we used Ridge, SVR, Decision Tree, Random Forest, AdaBoost, and XGBoost to establish the prediction model on SA reduction with these 17 features. Additionally, models were constructed using the top 10 and 17 radiomics features to compare the effect of DVH criteria and clinical features. The 17 integrated (10 radiomics, 4 clinical, 3 DVH criteria), top 17 radiomics, and top 10 radiomics features are shown in Table 2.
Because of the small sample size in this study, we used the leave-one-out method to validate the predictive performance of the model. Specifically, 51 patients were used for training, and the remaining one was tested, which repeated 52 times. MSE and R2 were used as the evaluation metrics. The performances are shown in Table 3. Based on the top 17 integrated features, the distribution of predicted values applying XGBoost versus the distribution of real values is shown in Fig. 5.
The experimental results showed that the prediction model of SA reduction based on XGBoost using top 17 integrated features (including top 10 radiomics features, 4 clinical, and 3 DVH criteria features) had the highest accuracy. In addition, we observed that applying the same algorithm, top 17 integrated features performed better than the other two feature groups.
4 Discussion
Most of studies used Pearson correlation method to pre-select features, but it ignored the information carried by the removed features. However, with XGBoost + SHAP, researchers can explore the impact of each feature, whether or not these features are ultimately used for modeling. In this study, PGs_V30, clinical stage, PGs_Dmin were ranked low in feature importance and were not used for modeling. But through XGBoost + SHAP, from Figs. 2c, e, we can explore the influence of these factors on SA reduction after RT, which enables researchers to obtain more comprehensive information.
Numerous studies showed that radiomics features can be used to predict xerostomia after RT with an accurate performance [7, 8, 10, 17, 34]. Similarly, in this study, XGBoost, AdaBoost, Random Forest, Decision Tree, SVR, and Ridge were applied to predict SA0f-30f reduction based on only radiomics features. All R2 from the models established by XGBoost, AdaBoost, Random Forest and Decision Tree were greater than 0.90, indicating that the models have high accuracy. It further supports that CT-based radiomics features are important in predicting the risk of xerostomia [19]. Previous studies showed that DVH criteria and clinical features also played an important role in predicting xerostomia after RT. To our knowledge, this is the first study to use radiomics, DVH criteria and clinical features to predict SA reduction. Adding DVH criteria and clinical features to radiomics features, the performance of all machine learning models was improved, which was similar to the findings of Sheikh et al. [48]. The results of applying different machine learning techniques with the same number of features showed that using integrated features (radiomics + DVH criteria + clinical) can improve the accuracy of the model than using only radiomics features. All in all, we believe that using radiomics, DVH criteria and clinical features to predict SA reduction after RT has huge potential value because this approach not only improves prediction performance but also increases the interpretability of the model. In this study, we used radiomics, DVH criteria and clinical features to predict SA reduction after RT. The best model is XGBoost with an MSE of 0.6994 and R2 of 0.9815.
In this study, we observed that PGs_Dmax, PGs_V15, and PGs_V45 were positively correlated with the severity of xerostomia. It was consistent with a large number of studies showing that the lower the mean dose of PGs, the lower the degree of xerostomia [22, 25,26,27]. There were fewer studies using PGs_Dmax, PGs_V15, and PGs_V45, but these characteristics reflect the dose to which the PGs are exposed. It was generally accepted that the high amount of radiation to which the PGs is exposed, increases the risk of radiation damage to PGs.
In this study, age also played an important role in predicting SA reduction after RT. It was widely accepted that age is an important prediction factor of xerostomia. However, there is still controversy regarding the relationship between age and xerostomia. Some studies suggested that age is negatively associated with xerostomia significantly [12, 25]. And some studies suggested that patients with large age have an increased risk of xerostomia due to the use of certain medications [49]. In this study, from Fig. 2c, it is showed that with the decrease of age, the severity of xerostomia increased. We believed that it may be related to the differences between the study samples. Meanwhile, we found that Hb had an important influence on the prediction of SA reduction. Currently, there is a lack of studies on the relationship between Hb level and xerostomia after RT in NPC patients. It is reported that an increase Hb concentration enhances the radiosensitivity of normal tissue such as skin and mucosa [50], which may increase PGs damage, leading to obvious SA reduction. Additionally, in this study, BMI was significant for predicting SA reduction after RT. Egestad et al. showed that patients with BMI ≥ 25 had more problems with xerostomia after RT than patients with BMI < 25 [51]. Sanguineti et al. also found that patients with a BMI > 30 had a significantly higher risk of xerostomia after RT [52]. The conclusions of the above studies are similar to those of this study. From Fig. 2c, it can be observed that BMI has a positive effect on SA reduction. Possibly, obese patients might experience relatively larger changes in anatomy, which might cause the salivary glands to receive higher doses [53].
In particular, our study has some limitations. First, we used radiomics features of PGs to predict SA reduction after RT. It is reported that about 60–70% of the stimulated SA originates from PGs, about 20% from the submandibular gland, and the rest from the other salivary glands [54]. Thus, we will add radiomics features of submandibular gland to the model to explore the role of submandibular gland. Second, this study was based on 52 NPC patients, and the small number of patients limited the generalizability of the models, so the validation on larger data sets will be needed in the future.
5 Conclusion
In this study, we used XGBoost + SHAP to feature selection. It can avoid multicollinearity while assisting the researcher to understand the impact of all features on the prediction target. We believed that it can be extended to studies predicting other diseases. Based on the radiomics features, adding DVH criteria and clinical features can effectively improve the accuracy of the model in predicting SA reduction. With the same number of features, using integrated features (radiomics + DVH criteria + clinical) can achieve better prediction performance than using only radiomics features. In this paper, the optimal combination of models for SA reduction is XGBoost + top 17 integrated features. The important DVH criteria and clinical features include WBC, PGs_Dmax, Age, PGs_V15, Hb, BMI and PGs_V45.
Data availability
The data that support the findings of this study are available from [The public scientific research data storage platform (www.researchdata.org.cn).], and the approval number is RDDB2018000256. But restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of [The public scientific research data storage platform].
References
Jensen AB, Hansen O, Jørgensen K, Bastholt L. Influence of late side-effects upon daily life after radiotherapy for laryngeal and pharyngeal cancer. Acta Oncol. 1994;33(5):487–91. https://doi.org/10.3109/02841869409083923.
De Leeuw V, Buffart LM, Heymans MW, Rietveld DH, Doornaert P, De Bree R, Buter J, Aaronson NK, Slotman BJ, Leemans CR. The course of health-related quality of life in head and neck cancer patients treated with chemoradiation: a prospective cohort study. Radiother Oncol. 2014;110(3):422–8. https://doi.org/10.1016/j.radonc.2014.01.002.
Lin A, Kim HM, Terrell JE, Dawson LA, Ship JA, Eisbruch A. Quality of life after parotid-sparing Imrt for head-and-neck cancer: a prospective longitudinal study. Int J Radiat Oncol Biol Phys. 2003;57(1):61–70. https://doi.org/10.1016/S0360-3016(03)00361-4.
Jellema AP, Slotman BJ, Doornaert P, Leemans CR, Langendijk JA. Impact of radiation-induced xerostomia on quality of life after primary radiotherapy among patients with head and neck cancer. Int J Radiat Oncol Biol Phys. 2007;69(3):751–60. https://doi.org/10.1016/j.ijrobp.2007.04.021.
Vissink A, Luijk PV, Langendijk JA, Coppes RP. Current ideas to reduce or salvage radiation damage to salivary glands. Oral Dis. 2015;21(1):e1–10. https://doi.org/10.1111/odi.12222.
Yovino S, Poppe M, Jabbour S, David V, Garofalo M, Pandya N, Alexander R, Hanna N, Regine WF. Intensity-modulated radiation therapy significantly improves acute gastrointestinal toxicity in pancreatic and ampullary cancers. Int J Radiat Oncol Biol Phys. 2011;79(1):158–62. https://doi.org/10.1016/j.ijrobp.2009.10.043.
Scalco E, Fiorino C, Cattaneo GM, Sanguineti G, Rizzo G. Texture analysis for the assessment of structural changes in parotid glands induced by radiotherapy. Radiother Oncol. 2013;109(3):384–7. https://doi.org/10.1016/j.radonc.2013.09.019.
Dijk L, Brouwer CL, Schaaf A, Burgerhof J, Steenbakkers R. Ct image biomarkers to improve patient-specific prediction of radiation-induced xerostomia and sticky saliva. Radiother Oncol. 2017;122(2):185–91. https://doi.org/10.1016/j.radonc.2016.07.007.
Van Dijk LV, Brouwer CL, Van Der Laan HP, Burgerhof JGM, Langendijk JA, Steenbakkers RJHM, Sijtsema NM. Geometric image biomarker changes of the parotid gland are associated with late xerostomia. Int J Radiat Oncol Biol Phys. 2017;99(5):1101–10. https://doi.org/10.1016/j.ijrobp.2017.08.003.
Pota M, Scalco E, Sanguineti G, Farneti A, Cattaneo GM, Rizzo G, Esposito M. Early prediction of radiotherapy-induced parotid shrinkage and toxicity based on ct radiomics and fuzzy classification. Artif Intell Med. 2017. https://doi.org/10.1016/j.artmed.2017.03.004.
Nardone V, Tini P, Nioche C, Mazzei MA, Carfagno T, Battaglia G, Pastina P, Grassi R, Sebaste L, Pirtoli L. Texture analysis as a predictor of radiation-induced xerostomia in head and neck patients undergoing Imrt. Radiol Med (Torino). 2018;123(6):415–23. https://doi.org/10.1007/s11547-017-0850-7.
Liu Y, Shi H, Huang S, Chen X, Zhou H, Chang H, Xia Y, Wang G, Yang X. Early prediction of acute xerostomia during radiation therapy for nasopharyngeal cancer based on delta radiomics from Ct images. Quant imaging Med Surg. 2019;9(7):1288.
Rosen BS, Hawkins PG, Polan DF, Balter JM, Brock KK, Kamp JD, Lockhart CM, Eisbruch A, Mierzwa ML, Ten Haken Rk. Early changes in serial Cbct-measured parotid gland biomarkers predict chronic xerostomia after head and neck radiotherapy. Int J Radiat Oncol Biol Phys. 2018;102(4):1319–29. https://doi.org/10.1016/j.ijrobp.2018.06.048.
Vernuccio Federica, Arnone Federica, Cannella Roberto, Verro Barbara, Comelli Albert, Agnello Francesco, Stefano Alessandro, et al. Diagnostic performance of qualitative and radiomics approach to parotid gland tumors: which is the added benefit of texture analysis? Br J Radiol. 2021;94(1128):20210340. https://doi.org/10.1259/bjr.20210340.
Zhang Z, Yang J, Ho A. A predictive model for distinguishing radiation necrosis from tumour progression after gamma knife radiosurgery based on radiomic features from Mr images. Eur Radiol. 2018;28(6):2255–63. https://doi.org/10.1007/s00330-017-5154-8.
Van Dijk LV, Thor M, Steenbakkers RJHM, Apte A, Zhai T, Borra R, Lee N, Langendijk JA, Deasy JO, Sijtsema NM. Oc-0180: parotid gland fat related magnetic resonance image biomarkers improve prediction of late xerostomia. Radiother Oncol. 2018;127:S95–6. https://doi.org/10.1016/S0167-8140(18)30490-0.
Liao YK, Chiu CC, Chiang WC, Chiou YR, Huang TC. radiomics features analysis of pet images in oropharyngeal and hypopharyngeal cancer. Medicine. 2019;98(18):e15446. https://doi.org/10.1097/MD.0000000000015446.
Wilkie JR, Mierzwa MM, Casper KA, Mayo CS, Rosen BS. Predicting late radiation-induced xerostomia with parotid gland pet biomarkers and dose metrics. Radiother Oncol. 2020. https://doi.org/10.1016/j.radonc.2020.03.037.
Wu H, Chen X, Yang X, Tao Y, Li XA. Early prediction of acute xerostomia during radiation therapy for head and neck cancer based on texture analysis of daily Ct. Int J Radiat Oncol Biol Phys. 2018;102(4):1308–18. https://doi.org/10.1016/j.ijrobp.2018.04.059.
Belli Maria Luisa, Scalco Elisa, Sanguineti Giuseppe, Fiorino Claudio, Broggi Sara, Dinapoli Nicola, Ricchetti Francesco, Valentini Vincenzo, Rizzo Giovanna, Cattaneo Giovanni Mauro. Early changes of parotid density and volume predict modifications at the end of therapy and intensity of acute xerostomia. Strahlentherapie Und Onkologie. 2014;190(11):1001–7. https://doi.org/10.1007/s00066-014-0669-2.
Zhang Yunyan, Dan Ou, Yajia Gu, He Xiayun, Peng Weijun. Evaluation of salivary gland function using diffusion-weighted magnetic resonance imaging for follow-up of radiation-induced xerostomia. Korean J Radiol. 2018;19(4):758–66. https://doi.org/10.3348/kjr.2018.19.4.758.
Eisbruch A, Kim HM, Terrell JE, Marsh LH, Dawson LA, Ship JA. Xerostomia and its predictors following parotid-sparing irradiation of head-and-neck cancer. Int J Radiat Oncol Biol Phys. 2001;50(3):695–704. https://doi.org/10.1016/S0360-3016(01)01512-7.
Deasy JO, Moiseenko V, Marks L, Chao K, Eisbruch A. Radiotherapy dose-volume effects on salivary gland function. Int J Radiat Oncol Biol Phys. 2010;76(3 Suppl):S58-63. https://doi.org/10.1016/j.ijrobp.2009.06.090.
Lee TF, Liou MH, Huang YJ, Chao PJ, Ting HM, Lee HY, Fang FM. Lasso Ntcp predictors for the incidence of xerostomia in patients with head and neck squamous cell carcinoma and nasopharyngeal carcinoma. Sci Rep. 2014;4(1):1–8. https://doi.org/10.1038/srep06217.
Teguh DN, Levendag PC, Ghidey W, Montfort KV, Kwa S. Risk model and nomogram for dysphagia and xerostomia prediction in head and neck cancer patients treated by radiotherapy and/or chemotherapy. Dysphagia. 2013;28(3):388–94. https://doi.org/10.1007/s00455-012-9445-6.
Miah AB, Gulliford SL, Clark CH, Bhide SA, Zaidi SH, Newbold KL, Harrington KJ, Nutting CM. Dose-response analysis of parotid gland function: what is the best measure of xerostomia? Radiother Oncol. 2013;106(3):341–5. https://doi.org/10.1016/j.radonc.2013.03.009.
Pan XB, Liu Y, Huang ST, Chen KH, Zhu XD. Predictors for improvement of xerostomia in nasopharyngeal carcinoma patients receiving intensity-modulated radiotherapy. Medicine. 2019;98(36):e17030. https://doi.org/10.1097/MD.0000000000017030.
Han P, Lakshminarayanan P, Jiang W, Shpitser I, Hui X, Sang HL, Cheng Z, Guo Y, Taylor RH, Siddiqui SA. Dose/volume histogram patterns in salivary gland subvolumes influence xerostomia injury and recovery. Sci Rep. 2019;9(1):1–9. https://doi.org/10.1038/s41598-019-40228-y.
Gabry HS, Florian B, Florian S, Henrik H, Mark B. Design and selection of machine learning methods using radiomics and dosiomics for normal tissue complication probability modeling of xerostomia. Front Oncol. 2018;8:35. https://doi.org/10.3389/fonc.2018.00035.
Montazeri A, Harirchi I, Vahdani M, Khaleghi F, Jarvandi S, Ebrahimi M, Haji-Mahmoodi M. The European organization for research and treatment of cancer quality of life questionnaire (Eortc Qlq-C30): translation and validation study of the iranian version. Support Care Cancer. 1999;7(6):400–6. https://doi.org/10.1007/s005200050300.
Cox JD, Stetz JA, Pajak TF. Toxicity criteria of the radiation therapy oncology group (Rtog) and the European Organization for Research and Treatment of Cancer (Eortc). Int J Radiat Oncol Biol Phys. 2015;31(5):1341–6. https://doi.org/10.1016/0360-3016(95)00060-c.
Gabry HS, Buettner F, Sterzing F, Hauswald H, Bangert M. Parotid gland mean dose as a xerostomia predictor in low-dose domains. Acta Oncologica. 2017. https://doi.org/10.1080/0284186X.2017.1324209.
Meirovitz A, Murdoch-Kinch CA, Schipper M, Pan C, Eisbruch A. Grading xerostomia by physicians or by patients after intensity-modulated radiotherapy of head-and-neck cancer. Int J Radiat Oncol Biol Phys. 2006;66(2):445–53. https://doi.org/10.1016/j.ijrobp.2006.05.002.
Trotti A, Colevas AD, Setser A, Basch E. Patient-reported outcomes and the evolution of adverse event reporting in oncology. J Clin Oncol. 2007;25(32):5121–7. https://doi.org/10.1200/JCO.2007.12.4784.
Murakami Y, Soyano T, Kozuka T, Ushijima M, Koizumi Y, Miyauchi H, Kaneko M, Nakano M, Kamima T, Hashimoto T, Yoshioka Y, Oguchi M. Dose-based radiomic analysis (dosiomics) for intensity modulated radiation therapy in patients with prostate cancer: correlation between planned dose distribution and biochemical failure. Int J Radiat Oncol Biol Phys. 2022;112(1):247–59. https://doi.org/10.1016/j.ijrobp.2021.07.1714.
Chair-Krishnapuram, B General, M General Chair-Shah, A Program Chair-Smola, C Program Chair-Aggarwal, D Program Chair-Shen, and R Program Chair-Rastogi. (2016) Proceedings of the 22nd Acm Sigkdd international conference on knowledge discovery and data mining. paper presented at the acm sigkdd international conference on knowledge discovery & data mining 2016.
Bianchi J, Ruellas A, Gonalves JR, Paniagua B, Cevidanes LHS. Osteoarthritis of the temporomandibular joint can be diagnosed earlier using biomarkers and machine learning. Sci Rep. 2020. https://doi.org/10.1038/s41598-020-64942-0.
Lundberg S, Lee SI. A unified approach to interpreting model predictions. In: Proceedings of the 31st international conference on neural information processing systems (NIPS’17). Curran Associates Inc., Red Hook, NY, USA; 2017, p. 4768–4777.
Hoerl AE, Kennard RW. Ridge regression: biased estimation for nonorthogonal problems. Technometrics. 2012;12:55–67. https://doi.org/10.2307/1271436.
Wei Wei, Wang Ke, Liu Zhenyu, Tian Kaibing, Wang Liang, Jiang Du, Ma Junpeng, Wang Shuo, Li Longfei, Zhao Rui, Cui Luo, Zhen Wu, Tian Jie. Radiomic signature: a novel magnetic resonance imaging-based prognostic biomarker in patients with skull base chordoma. Radiother Oncol. 2019;141:239–46. https://doi.org/10.1016/j.radonc.2019.10.002.
Smola AJ, Schölkopf B. A tutorial on support vector regression. Stat Comput. 2004;14(3):199–222. https://doi.org/10.1023/B:STCO.0000035301.49549.88.
Drucker H. Improving regressors using boosting techniques. ICML. 1997;1997:107–15.
Sakai M, Nakano H, Kawahara D, Tanabe S, Utsunomiya S. Detecting Mlc modeling errors using radiomics-based machine learning in patient: pecific Qa with an Epid for intensity odulated radiation therapy. Med Phys. 2020;48(3):991–1002. https://doi.org/10.1002/mp.14699.
Liaw A, Wiener M. Classification and regression by randomforest. R news. 2002;2(3):18–22.
Homayounieh F, Saini S, Mostafavi L, Khera RD, Kalra MK. Accuracy of radiomics for differentiating diffuse liver diseases on non-contrast Ct. Int J Comput Assist Radiol Surg. 2020;15(9):1–10. https://doi.org/10.1007/s11548-020-02212-0.
Solomatine, Dimitri P, and Durga L Shrestha. Adaboost. Rt: A boosting algorithm for regression problems. Paper presented at the 2004 IEEE International Joint Conference on Neural Networks (IEEE Cat. No. 04CH37541); 2004: 1163–1168.
Thongkam, J., G. Xu, and Y. Zhang. 2008 Adaboost algorithm with random forests for predicting breast cancer survivability. Paper presented at the IEEE international joint conference on neural networks; 2008.
Sheikh K, Sang HL, Zhi C, Lakshminarayanan P, Lee J. Predicting acute radiation induced xerostomia in head and neck cancer using Mr and Ct radiomics of parotid and submandibular glands. Radiat Oncol. 2019;14(1):1–11. https://doi.org/10.1186/s13014-019-1339-4.
Beetz Ivo, Schilstra Cornelis, van der Schaaf Arjen, van den Heuvel Edwin R, Doornaert Patricia, van Luijk Peter, Vissink Arjan, van der Bernard FAM, Laan Charles R, Leemans Henk P, Bijl Miranda E.M.C., Christianen Roel J.H.M., Steenbakkers Johannes A, Langendijk. Ntcp models for patient-rated xerostomia and sticky saliva after treatment with intensity modulated radiotherapy for head and neck cancer: the role of dosimetric and clinical factors. Radiother Oncol. 2012;105(1):101–6. https://doi.org/10.1016/j.radonc.2012.03.004.
Elsaid A, Farouk M. Significance of anemia and role of erythropoietin in radiation induced mucositis in head and neck cancer patients. Int J Radiat Oncol Biol Phys. 2001;51(3):368. https://doi.org/10.1016/S0360-3016(01)02504-4.
Egestad H, Nieder C. Differences in quality of life in obese and normal weight head and neck cancer patients undergoing radiation therapy. Support Care Cancer. 2015;23(4):1081–90. https://doi.org/10.1007/s00520-014-2463-1.
Sanguineti G, Ricchetti F, Binbin Wu, McNutt T, Fiorino C. Parotid gland shrinkage during imrt predicts the time to xerostomia resolution. Radiat Oncol. 2015;10(1):1–6. https://doi.org/10.1186/s13014-015-0331-x.
You SH, Kim SY, Lee CG, Keum KC, Kim JH, Lee IJ, Kim YB, Koom WS, Cho J, Kim SK. Is there a clinical benefit to adaptive planning during tomotherapy in patients with head and neck cancer at risk for xerostomia? Am J Clin Oncol. 2012;35(3):261–6. https://doi.org/10.1097/COC.0b013e31820dc092.
Dawes C, Wood CM. The contribution of oral minor mucous gland secretions to the volume of whole saliva in man. Arch Oral Biol. 1973;18(3):337–42. https://doi.org/10.1016/0003-9969(73)90156-8.
Acknowledgements
The authors would like to thank Hongyu Shi, Yijun Hua, Haojiang Li, Yunfei Xia, for their valuable comments.
Funding
This work was supported by the National Natural Science Foundation of China (No. 82171906); the Basic and Applied Basic Research Foundation of Guangdong Province (2021A1515220140); the Youth Innovation Project of Sun Yat-sen University Cancer Center (QNYCPY32); the Student's Platform for Innovation and Entrepreneurship Training Program (202213902102, S202213902029, S202113902030); the Guangdong Medical Science and Technology Research Fund Project (A2020516); and the National Key Projects of Research and Development of China (2016YFC0904600).
Author information
Authors and Affiliations
Contributions
All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by XY, SH, LZ and WZ. The first draft of the manuscript was written by LZ and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.
Corresponding authors
Ethics declarations
Ethics approval and consent to participate
This project was approved by the Ethical Committee of Sun Yat-Sen University Cancer Center and informed consent was waived by the committee. The study was in accordance with the guidelines and regulations of Ethical Committee of Sun Yat-Sen University Cancer Center and WMA declaration of Helsinki.
Competing interests
The authors have no competing interests to declare.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Zhou, L., Zheng, W., Huang, S. et al. Integrated radiomics, dose-volume histogram criteria and clinical features for early prediction of saliva amount reduction after radiotherapy in nasopharyngeal cancer patients. Discov Onc 13, 145 (2022). https://doi.org/10.1007/s12672-022-00606-x
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s12672-022-00606-x