A predictive model of radiation-related fibrosis based on the radiomic features of magnetic resonance imaging and computed tomography.

BACKGROUND: To establish a predictive model for the fibrotic level of neck muscles after radiotherapy by using radiomic features extracted from the magnetic resonance imaging (MRI) before and after radiotherapy and planning computed tomography (CT) in nasopharyngeal carcinoma patients.
METHODS: A total of one hundred and eighty-six patients were finally enrolled in this study. According to the specific standard, all patients were divided into three different fibrosis groups. Regions of interests (ROI), including sternocleidomastoids (SCMs), trapezius (T), levator scapulae (LS), and scalenus muscles (S), were delineated manually and used for features extraction on IBEX. XGBoost, a machine learning algorithm, was used for the establishment of the prediction model. First, the patients were divided into training cohort (80%) and testing cohort (20%) randomly. Then the image features of CT or delta changes calculated from pre- and post-radiotherapy MRI images on each cohort constituted training and testing datasets. Then, based on the training dataset, a well-trained prediction model was produced. We used five-fold cross-validation to validate the predictive models. Afterward, the model performance was assessed on the 'testing' set and reported in terms of area under the receiver operating characteristic curve (AUC) under five scenarios: (I) only T1 sequence, (II) only T2 sequence, (III) only T1 post-contrast (T1 + C) sequence, (IV) Combination of all MRI sequences, (V) only CT.
RESULTS: Most of the patients enrolled are male (73.1%), mean age was 47 years, receiving concurrent chemo-radiotherapy as the primary treatment (90.9%). By the end of the final follow-up, most of the patients were rated as mild fibrosis (60.8%). We found the prediction model based on the CT image features outperform all MRI features with an AUC of 0.69 and accuracy of 0.65. Contrarily, the model based on features from all MRI sequence showed lower AUC less than 0.5 and lower accuracy less than 0.6.
CONCLUSIONS: The prediction model based on CT radiomics features has better performance in the prediction of the grade of post-radiotherapy neck fibrosis. This might help guide radiotherapy treatment planning to achieve a better quality of life. 2020 Translational Cancer Research. All rights reserved.

Introduction

Nasopharyngeal carcinoma is relatively rare worldwide. In 2018, 129,079 new cases of this malignancy were diagnosed, accounting for 0.7% of all cancers (1). Still, nasopharyngeal carcinomas are commonly diagnosed in Asia, especially in China [the crude incidence rate was 3.26/100,000 (2)]. Radiotherapy remains the standard treatment for these generally radiosensitive tumors and current estimates of 5-year overall survival after radiotherapy range from 66% to 83% (3). Given these promising survival rates, oncologists increasingly have focused on the quality of life of their patients. Radiation-related fibrosis is a typical late-onset complication of radiotherapy (4). This sequela may not arise until a few years after the end of the treatment and may progress or deteriorate further over time (4,5). Fibrosis can impair the functions of muscles in the head and neck and may thus restrict the opening of the mouth and jaw, motion in the shoulders, and rotation of the neck. Also, neck fibrosis may also cause cranial nerve palsy by compressing the hypoglossal nerve (6,7). All of these restrictions can interfere with eating, speaking, driving, self-care, and employment (5), severely impairing the quality of life. Radiation fibrosis is a multi-stage development process regulated by a variety of molecules, so it is difficult to design drugs that work at all stages. The current treatment strategies are mainly focused on limiting the aggravation of fibrosis, including topical emulsions (8), antioxidant therapies (9), hyperbaric oxygen therapy (10), adipose-derived stem cells (11) and some other therapies directly inhibiting the inflammatory mediators (12). The acupuncture and moxibustion therapy of traditional Chinese medicine also show a certain curative effect (13). However, the treatment options for radiation-related fibrosis are limited, and their therapeutic effects cannot sufficiently reverse the evolution and progression of fibrosis. Therefore, it is crucial to identify patients potentially at a high risk of fibrosis, as this will allow the application of preventive interventions to ensure optimal function and quality of life or minimize the side effects of treatment.

Fibrosis is characterized by an increase in tissue stiffness (i.e., loss of compliance), which can be detected by palpation. Accordingly, most studies have used hand palpation and clinician-based rating scales of fibrosis, including the Medical Research Council (MRC) (14), European Organization for Research and Treatment of Cancer/Radiation Therapy Oncology Group (EORTC/RTOG) (15), and Late Effects in Normal Tissues/Subjective, Objective, Management, and Analytic (LENT/SOMA) scoring systems (16). However, these scales are inevitably subjective, semiquantitative (17), and prone to interobserver error. Other studies have applied quantitative mechanical methods (18), quantitative electrical methods (19,20), ultrasound shear wave elastography (21), and magnetic resonance imaging (MRI) (22,23) to various parts of the body for fibrosis assessment. In the neck, tissue fibrosis may affect multiple tissue layers with considerable overlap. Currently, all of these changes cannot be detected using a single measurement technique, and even invasive biopsy is limited due to the ability to sample only specific microscopic points.

The field of radiomics is not based on information from images or single pathological tissue layers. Accordingly, this field differs from the traditional practice of subjecting medical images solely to visual interpretation (24). Radiomics exhibits tremendous promise as a comprehensive method of three-dimensional examination that enables the noninvasive profiling of multiple tissue layers (25). Nowadays, more and more researches begin to focus on the functional evaluation by using radiomics. A previous exploratory study succeeding in finding the relationship between the radiation dose to the masseter and the medial pterygoid and the variance of the MRI intensity of the radiation-induced trismus described by the radiomic textures (26). At the same time, a number of studies have shown the relationship between the image features and radiation-induced xerostomia and the integration of image features into the predictive model may improve the risk stratification of xerostomia (27-29). MRI is used widely in clinical workups for the pretreatment diagnosis and staging of nasopharyngeal carcinoma and as a conventional routine follow-up method. And every patient would perform planning CT scans before radiotherapy. In this study, therefore, we aimed to grope for if the radiomic features extracted from many types of MRI scans or CT could build a predictive model of radiation-related fibrosis.

Methods

The workflow of this study is depicted in . Specific contents and details in the process are described below.

Figure 1

A workflow to produce a predictive model. These steps include gathering a set of patient images, segmentation of the region of interest (ROI) including four muscle, extracting a set of radiomics features from these ROIs, generating a predictive model and then perform the statistical analysis.

Patient cohort

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by Ethic Committee of the Hunan Cancer Hospital (No. 07 of 2020 Scientific Research Quick Review). The study is a retrospective study that examines the fait accompli of the past. Before the retrospective study, all the private information of patients will be anonymized, no direct contact with patients, no privacy of patients, the results of the study are only used for medical research, there is no risk to the patients included in the study. Based on the above, we applied for exemption from the informed consent of the subject and approved by the Ethic Committee.

For this study, we initially analyzed all MRI and CT data of 749 patients from the same institution in 2015, who received chemoradiation in 6–7 weeks. All MR images were obtained from a 1.5-T MRI system (GE sigma CV/i) during routine clinical practice. All photos were axial scans with the field of view of 30 cm, slice thickness of 5 mm, and slice spacing of 1 mm. CT images were acquired on GE LightSpeed RT (GEHW) with a peak tube voltage of 120 kVp and exposure of 200mAs. Images had 512×512 pixels, FOV of 550.0 mm × 550.0 mm, and a slice thickness of 5 mm.

The final analysis included the images, clinical factors, and outcome data of the patients who met the following inclusion criteria: (I) initial treatment for pathologically confirmed nasopharyngeal carcinoma at a single institution in 2015; (II) no history of radiotherapy of neck and head because of other disease; (III) treated with static intensity-modulated radiation therapy (s-IMRT), step and shoot; (IV) availability of MR images collected at two points (before radiotherapy and up to 6 months post-treatment), including T1-weighted (T1), T1 post-contrast (T1 + C), and T2-weighted (T2) scans; (V) availability of planning CT images. Patients were excluded when they met the following criteria: (I) patients who could not be contacted at the follow-up time; (II) unavailability of MR images at each point or planning CT images; (III) poor image quality which is not sufficient for diagnosis and analysis. The specific screening process is shown in .

Figure 2

Flow diagram for patient selection in our study.

Clinical factors

The patient’s clinical parameters were obtained through a retrospective review and are presented systematically in . The median follow-up duration was 18 months (range, 11–22 months). The patients were asked about the presence of four symptoms: (I) discomfort in the neck (self-rated on a scale of 0–10), (II) experience of late facial edema after treatment (yes or no, if yes, how long did it regress), (III) experience of upper limb pain after treatment (yes or no, if yes, the severity and is there any treatment), and (IV) restricted head rotation during activities of daily living (yes or no, if yes, the degree of limitation and how long does it last). Based on these symptoms and physical signs, all patients were divided into three groups: mild, moderate, and severe fibrosis. Specific standards are listed in .

Table 1

Patient’s clinical and dosimetric parameters

Clinical factors	N=186	Training cohort					Testing cohort
		Mild	Moderate	Severe	P		Mild	Moderate	Severe	P
SEX					0.972					0.645
Male	136	67	24	18			17	6	4
Female	50	23	12	4			6	3	2
AGE (years)					0.005					0.273
≥65	8	1	2	3			1	1	0
<65	178	89	34	19			22	8	6
T stage					0.935					0.852
T1	19	10	3	2			2	1	1
T2	59	28	14	5			7	4	1
T3	60	30	10	8			7	2	3
T4	48	22	9	7			7	2	1
N stage					0.398					0.714
N0	4	1	1	1			1	0	0
N1	28	13	6	3			3	1	2
N2	112	58	20	11			14	5	4
N3	42	18	9	7			5	3	0
M stage					0.03					0.524
M0	178	88	36	18			22	9	5
M1	8	2	0	4			1	0	1
Clinical stage					0.222					0.672
I	2	0	1	0			0	1	0
II	11	6	2	1			2	0	0
III	91	45	18	10			11	4	3
IVa	74	37	15	7			9	4	2
IVb	8	2	0	4			1	0	1
Concurrent chemoradiotherapy					0.641					0.687
No	17	9	4	1			2	1	0
Yes	169	81	32	21			21	8	6

The 7th American Joint Committee on Cancer (AJCC) TNM staging manual was used to stage the patients.

Table 2

Categorization of 186 patients according to muscle fibrosis in the neck after treatment

Level	Conditions (meet any of the following)
Mild	(I) Rating: 0–4 points
	(II) No facial edema
	(III) No upper limb pains
	(IV) Neck activity is unrestricted
Moderate	(I) Rating: 4–6 points
	(II) Late facial edema (regress within 3 months)
	(III) Mild upper limb pain (no treatment)
	(IV) Neck activity is slightly limited
Severe	(I) Rating: 7–10 points
	(II) Late facial edema, lasting more than 3 months
	(III) Severe upper limb pain, or needing medical intervention
	(IV) Neck activity is significantly limited or duration >6 months

Categorization into a specific group required the patient to meet any of the above criteria.

Lesion segmentation

Lesion segmentation was performed using The Imaging Biomarker Explorer (IBEX) software package, version 1.0 (30). An experienced clinical oncologist used the IBEX software to contour the ROIs manually on each type of MRI sequence (T1, T1 + C, and T2) and CT for each patient. Later, the contoured ROI would be reviewed by another experienced radiologist. It has been reported that muscle fibrosis develops before skin fibrosis (31). Based on this theory, the ROI was drawn to include sternocleidomastoids (SCMs), trapezius (T), levator scapulae (LS), and scalenus muscles (S). The ROI was divided horizontally into two parts based on the level of the cricoid cartilage, which received a different radiation dose during radiotherapy. It is worth reminding that we contoured four ROIs above the cricoid cartilage on CT because of the loss of part images.

Extraction and calculation of radiomic features

IBEX was used to extract image features from pre- and post-radiotherapy MRI images or planning CT images. Categorization according to specific standards () was used to label the image features like three different groups. As shown in , the textural features calculated in this study can be organized into 10 categories. Two processing methods were applied, including Resample Voxel Size and Butterworth Smooth (32), which lessened the image noise. Assigning different parameters to every textural feature, and a total of 190 texture features were acquired for each sequence of the MRI at every time points and 1,767 for CT. The magnitude of change of each feature was computed as follows: delta change = (post − pre)/pre, being used for MRI feature modeling, where post and pre correspond to the measurements after and before radiation therapy, respectively. Changes from the first time to the second time in radiomics features, called delta-radiomics, have been proved to improved model for predicting the prognosis of patients combined with clinical factors and radiomic characteristics (33). On the meanwhile, features acquired on CT were directly used for modeling on the next step.

Table 3

The features used in this study and related preprocessing methods

Category	Feature	Preprocess
Gradient Orient Histogram	Inter Quartile Range	Resample Voxel Size
	Kurtosis
	Mean Absolute Deviation
	Median Absolute Deviation
	Percentile
	Percentile Area
	Quantile
	Range
	Skewness
Gray Level Co-occurrence Matrix 25	Auto Correlation	Resample Voxel Size
	Cluster Prominence
	Cluster Shade
	Cluster Tendency
	Contrast
	Correlation
	Difference Entropy
	Dissimilarity
	Energy
	Entropy
	Homogeneity
	Homogeneity2
	InformationMeasureCorr1
	InformationMeasureCorr2
	Inverse Diff Moment Norm
	Inverse Diff Norm
	Inverse Variance
	Max Probability
	Sum Average
	Sum Entropy
	Sum Variance
	Variance
Gray Level Co-occurrence Matrix 3	Auto Correlation	Resample Voxel Size
	Cluster Prominence
	Cluster Shade
	Cluster Tendency
	Contrast
	Correlation
	Difference Entropy
	Dissimilarity
	Energy
	Entropy
	Homogeneity
	Homogeneity2
	InformationMeasureCorr1
	InformationMeasureCorr2
	Inverse Diff Moment Norm
	Inverse Diff Norm
	Inverse Variance
	Max Probability
	Sum Average
	Sum Entropy
	Sum Variance
	Variance
Gray Level Run Length Matrix25	Gray Level Nonuniformity	Resample Voxel Size, Butterworth Smooth
	High Gray Level Run Emphasis
	Long Run Emphasis
	Long Run High Gray Level Emphasis
	Long Run Low Gray Level Emphasis
	Low Gray Level Run Emphasis
	Run Length Nonuniformity
	Run Percentage
	Short Run Emphasis
	Short Run High Gray Level Emphasis
	Short Run Low Gray Level Emphasis
Intensity Direct	Energy	Resample Voxel Size
	Energy Norm
	Global Entropy
	Global Max
	Global Mean
	Global Median
	Global Min
	Global Std
	Global Uniformity
	Inter Quartile Range
	Kurtosis
	Local Entropy Max
	Local Entropy Mean
	Local Entropy Median
	Local Entropy Min
	Local Entropy Std
	Local Range Max
	Local Range Mean
	Local Range Median
	Local Range Min
	Local Range Std
	Local Std Max
	Local Std Mean
	Local Std Median
	Local Std Min
	Local Std Std
	Mean Absolute Deviation
	Median Absolute Deviation
	Percentile
	Quantile
	Range
	Root Mean Square
	Skewness
Intensity Histogram	Inter Quartile Range	Resample Voxel Size, Butterworth Smooth
	Kurtosis
	Mean Absolute Deviation
	Median Absolute Deviation
	Percentile
	Percentile Area
	Quantile
	Range
	Skewness
Neighbor Intensity Difference 25	Busyness	Resample Voxel Size, Butterworth Smooth
	Coarseness
	Complexity
	Contrast
	Texture Strength
Neighbor Intensity Difference 3	Busyness	Resample Voxel Size, Butterworth Smooth
	Coarseness
	Complexity
	Contrast
	Texture Strength
Shape	Compactness1	–
	Compactness2
	Convex
	Convex Hull Volume
	Convex Hull Volume 3D
	Mass
	Max3D Diameter
	Mean Breadth
	Number of Objects
	Number of Voxel
	Orientation
	Roundness
	Spherical Disproportion
	Sphericity
	Surface Area
	Surface Area Density
	Volume
	Voxel Size
Intensity Histogram Gaussian Fit	Gaussian Amplitude	Resample Voxel Size
	Gaussian Area
	Gaussian Mean
	Gaussian Std
	Hist Area
	Number of Gaussian

Feature modeling

In reality, many features have high noise and may lead to overfitting or classification errors in feature modeling. Not all image features can be conductive to grade the severity of neck fibrosis after radiotherapy. Features were selected for the prediction performance in terms of the AUC (>0.6). Meanwhile, XGBoost, a Gradient Tree Boosting regularization form (34), could identify a subset of essential features to avoid feature redundancy for feature modeling in the course of calculation. Previous studies have reported that XGBoost showed lower test error rate and the larger AUC in comparison with logistic regression analysis and other machine learning approaches, including decision tree, random forest, and support vector machine (35). It leads us to propose using this method for the establishment of the prediction model of radiation-related fibrosis. First, the subjects included were divided into five folds which was consistent with the proportion of the three fibrotic groups in the overall cohort randomly. Then four folds constituted the training cohort, and the remaining one constituted the testing cohort. The modeling was done based on the selected image features delta changes on the MRI images that were calculated in the previous step or the features extracted from CT on the cohort. Next, XGBoost was used to generate a well-trained prediction model based on the training image dataset. Though a lack of independent external validation, we do perform five-fold cross-validation (internal validation), which is identified to be an effective way to build patient-specific predictions without bias (36). Finally, we compared the prediction performance under five scenarios: (I) only the T1 sequence; (II) only the T2 sequence; (III) only the T1 + C sequence; (IV) combination all the MRI sequences; (V) only the CT. The performance of the predictive model was then assessed using the testing image set, and the results are reported in terms of the mean AUC.

Results

Patients cohort and selected features

As shown in , lastly, a total of 186 patients were finally enrolled for further analysis. The patients include was divided into training cohort (80.0%) and testing cohort (20%) randomly. The patient sample was predominantly male (73.1%), with a mean age of 47 years. The majority had locally advanced disease (93.0%) and had received concurrent chemoradiotherapy as the primary treatment (90.9%). By the end of the final follow-up, most patients were classified as having mild fibrosis (60.8%). There are no significant differences among the fibrosis groups for all the factors except age (P=0.005) and metastasis stage (P=0.03) of the patients in the training cohort. Detailed clinical parameters are depicted in . As shown in , fibrosis degree has nothing to do with the lymph node stage. We performed spearman relativity analysis between these two variables (P=0.613, rs=0.029).

Figure 3

A hotspot map of the relationship between lymph node stage and the degree of fibrosis.

A total of 139 textures from T1 sequence, 138 features from T2 sequence and 157 features from T1 + C sequence were used for feature modeling and 749 features for CT. As shown in , extracted features include first-order, second-order and higher-order characteristic, which can be further grouped into ten categories. Shape features, Neighbor Intensity Difference features, Intensity Histogram Gaussian Fit features, Intensity histogram features, Intensity Direct features (37), Gray Level Co-occurrence Matrix features, Gray Level Run Length Matrix features (38), and Gradient Orient Histogram features (39) were enrolled in this study.

The radiomics signatures based on MRI images or CT images performed well in predicting the degree of fibrosis after radiotherapy. Values of AUC, accuracy, sensitivity and specificity of the five scenarios are as shown in . When predicting three different fibrosis groups, we found the model based on the CT image features showed better performance, with an AUC of 0.69 and accuracy of 0.65 compared to that models of T1, T2, T1 + C, and combine all three sequences show lower AUC (all is 0.49) and lower accuracy (0.56, 0.55, 0.57, 0.58).

Table 4

The performance of the predictive model using the XGBoost

Modality	AUC	Accuracy	Degree	Sensitivity (%)	Specificity (%)
T1	–	–	(I) Mild	81.39	20.10
	0.49	0.56	(II) Moderate	13.33	83.87
	–	–	(III) Severe	3.33	96.20
T2	–	–	(I) Mild	84.04	6.71
	0.49	0.55	(II) Moderate	4.00	88.45
	–	–	(III) Severe	3.00	96.60
T1 + C	–	–	(I) Mild	81.35	19.29
	0.49	0.57	(II) Moderate	17.33	83.34
	–	–	(III) Severe	0.20	97.73
T1 + T2 + T1 + C		–	(I) Mild	85.09	14.52
	0.49	0.58	(II) Moderate	13.33	88.93
	–	–	(III) Severe	3.00	96.58
CT	–	–	(I) Mild	98.74	2.09
	0.69	0.65	(II) Moderate	2.15	98.66
	–	–	(III) Severe	0.00	100.00

AUC, area under the curve; T1, T1-weighted scans; T1 + C, T1 post-contrast cans; T2, T2-weighted scans; CT, computed tomography.

Discussion

In this study, we developed predictive models of radiation-related fibrosis based on the radiomic features of MRI and CT. These features exhibited predictive power to a certain extent. To our best knowledge, our group was the first to use the radiomic features of MRI and CT to predict the grade of radiation-related fibrosis on the neck. Additionally, we proposed a more practical standard for fibrosis level, including symptoms and physical signs.

Notably, the features extracted from CT outperformed all other feature changes from commonly used MRI scans, including T1, T2, and T1 + C, in terms of AUC values and accuracy. Compared with CT, MRI has the advantage of showing soft tissue lesions better and providing muscle-specific measurements. Researchers have previously identified a close relationship between MRI features and the severity of radiation-induced fibrosis (40). However, we did find that image features from CT have higher accuracy in predicting the degree of fibrosis after radiotherapy. Despite we only contoured four ROIs on CT images compared to eight ROIs on MRI, more features were extracted from CT. Besides, the type of acquisition noise, enhancement status, and image reconstruction algorithm have different effects on MRI imaging characteristics, especially in a retrospective study (41). To our knowledge, no standardized MRI method has been developed for the purpose of head and neck scanning or radiomics. Therefore, the effects of various MRI factors on the image data cannot yet be avoided. On the other hand, despite no existing studies have correlated clinically rated neck fibrosis with CT findings, previous studies have found textural image features extracted from CT are highly correlated with the severity of pulmonary fibrosis (42) and could discriminate between patients with and those without radiation pneumonitis (43).Moreover, the previous study has used CT texture changes for distinguishing radiation-induced fibrosis from tumor recurrence for lung cancer (44).

Despite the advantages of this approach, our research had some limitations. First, standard protocols for fibrosis grading included patient self-ratings, which inevitably led to subject bias. Second, the current radiomic analysis protocol involves complex computational steps with frequent human interactions and a potentially time-consuming analytical process and may be challenging to include in daily clinical practice. Currently, this post-processing process includes multiple steps and calculations and requires approximately 60 minutes per patient (45). But we believe that as technological innovation and the optimization of the algorithm, it will be less time-consuming in the future. Moreover, our study did not distinguish between radiation-induced fibrosis and residual or recurrent tumor. The residual tumor remained after treatment in nearly 7–13% of nasopharyngeal carcinoma cases (46). A future prospective study should incorporate data from dynamic contrast-enhanced MRI, the primary choice for the diagnosis of neck fibrosis after radiotherapy for nasopharyngeal carcinoma, given its ability to identify tumor residue, recurrence, and fibrosis (47). Finally, our analysis was based on a retrospective design and data from a single center. Our model requires validation through a prospective multicenter trial with a larger study cohort.

Conclusions

In conclusion, we constructed a predictive, non-invasive, inexpensive, and highly patient-specific model of radiation-related fibrosis that does not affect existing clinical activities. This model requires further optimization, but it does contribute to the decision-making of the radiation treatment. Oncologists may use this model to compare the potential effects of different therapeutic regimens on the grade of radiation-related fibrosis and could thus individually tailor treatments to minimize the side reaction by adjusting the dose prescription on the neck. Furthermore, our model may inform studies of radiation-related injuries in other body regions. In the future, we aim to develop the model further to enable direct predictions of the effects of radiation-related fibrosis on the quality of life.