Survival rate prediction of nasopharyngeal carcinoma patients based on MRI and gene expression using a deep neural network

Abstract

To achieve a better treatment regimen and follow‐up assessment design for intensity‐modulated radiotherapy (IMRT)‐treated nasopharyngeal carcinoma (NPC) patients, an accurate progression‐free survival (PFS) time prediction algorithm is needed. We propose developing a PFS prediction model of NPC patients after IMRT treatment using a deep learning method and comparing that with the traditional texture analysis method. One hundred and fifty‐one NPC patients were included in this retrospective study. T1‐weighted, proton density and dynamic contrast‐enhanced magnetic resonance (MR) images were acquired. The expression level of five genes (HIF‐1α, EGFR, PTEN, Ki‐67, and VEGF) and infection of Epstein–Barr (EB) virus were tested. A residual network was trained to predict PFS from MR images. The output as well as patient characteristics were combined using a linear regression model to provide a final PFS prediction. The prediction accuracy was compared with that of the traditional texture analysis method. A regression model combining the deep learning output with HIF‐1α expression and Epstein–Barr infection provides the best PFS prediction accuracy (Spearman correlation R ² = 0.53; Harrell's C‐index = 0.82; receiver operative curve [ROC] analysis area under the curve [AUC] = 0.88; log‐rank test hazard ratio [HR] = 8.45), higher than a regression model combining texture analysis with HIF‐1α expression (Spearman correlation R ² = 0.14; Harrell's C‐index =0.68; ROC analysis AUC = 0.76; log‐rank test HR = 2.85). The deep learning method does not require a manually drawn tumor region of interest. MR image processing using deep learning combined with patient characteristics can provide accurate PFS prediction for nasopharyngeal carcinoma patients and does not rely on specific kernels or tumor regions of interest, which is needed for the texture analysis method.

Progression‐free survival (PFS) time prediction of nasopharyngeal carcinoma (NPC) patients can be achieved by analyzing multi‐contrast MR images. A deep residual neural network was trained to predict PFS for NPC patients from T1‐weighted, proton‐density, and dynamic contrast‐enhanced MR images. Deep learning‐based PFS prediction showed better performance than texture analysis‐based PFS prediction.

Abbreviations

AJCC/UICC

American Joint Committee on Cancer/Union for International Cancer Control

CI

confidence interval

DCE

dynamic contrast‐enhanced MRI

DL

deep learning

EBV

Epstein–Barr virus

EGFR

epidermal growth factor receptor

HIF‐1α

hypoxia inducible factor‐1‐alpha

HR

hazard ratio

IMRT

intensity‐modulated radiotherapy

NPC

nasopharyngeal carcinoma

PD

proton density MRI

PFS

Progression‐free survival

PTEN

Phosphatase and tensin homolog

T1w

T1‐weighted MRI

TA

texture analysis

VEGF

Vascular Endothelial Growth Factor

1. INTRODUCTION

Nasopharyngeal carcinoma (NPC) is an epithelial carcinoma arising from the nasopharyngeal mucosa lining with an unbalanced geographical global distribution: 70% of NPC patients are in East or Southeast Asia. ¹ , ² Intensity‐modulated radiotherapy (IMRT) is the primary NPC treatment method because of the sensitivity of NPC to radiotherapy. The combination of IMRT with different therapeutics may improve patient survival rates. ¹ For example, evofosfamide targets the hypoxia tumor cells and leads to an enhanced radiotherapeutic response; ³ nimotuzumab targets the epidermal growth factor receptor (EGFR) and leads to inhibition of downstream EGFR signaling. ⁴ The combination of chemotherapy with radiotherapy may further improve the survival rate and help with locoregional control of nonmetastatic NPC. ⁵ Therefore, it has become the most commonly used NPC treatment method.

While the development of novel therapeutics and the optimization of chemotherapy strategies (induction, concurrent, and adjuvant) reflect a promising treatment direction for NPC, the treatment response for NPC patients is still not satisfactory, especially for advanced NPC patients. The overall 5‐year survival rate for advanced NPC patients is between 38% and 63%, while the overall 5‐year survival rate for NPC patients treated with IMRT is between 59% and 83%. ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ An accurate progression‐free survival (PFS) prediction method is needed for a better treatment regimen and follow‐up assessment design.

Progression‐free survival prediction can be performed based on various indexes, such as blood markers, gene and protein expression levels, and image features. ¹¹ , ¹² , ¹³ With the fast development of MRI techniques, the potential of using dynamic contrast enhancement (DCE) and T1‐weighted magnetic resonance (MR) images for PFS prediction has been evaluated in various pre‐clinical and clinical studies, as they can detect the morphology, flow, and permeability differences of NPCs. ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ Apart from MR images, gene expression levels detected by immunohistochemistry from electro‐nasopharyngeal fiberoptic specimens can reflect hypoxic and proliferative gene expression information in NPC tumors. Therefore, they may also contribute to PFS prediction. Hypoxia‐inducible factor 1 alpha (HIF‐1α) is an essential regulator of tumor cell adaptation to hypoxia and regulates the angiogenesis, energy metabolism, invasion/metastasis, and radiotherapy tolerance of tumors. ¹⁸ Epidermal growth factor receptor (EGFR), a proliferation biomarker, is associated with advanced stage and poor prognosis in NPC and can also induce HIF‐1α expression in a hypoxic microenvironment. ¹⁹ Ki‐67 is an indicator of cell proliferation and tumor aggressiveness. ²⁰ The phosphatase and tensin homolog (PTEN) gene controls the generation of two enzymes that act as tumor suppressors by regulating cell division. ²¹ Vascular endothelial growth factor (VEGF) encodes the heparin‐binding protein, which is essential for pathological angiogenesis. ²² The Epstein–Barr virus (EBV) is strongly associated with the progression of NPC. ²³ Recent genetic studies have shown that EBV infection may transform pre‐invasive nasopharyngeal epithelial cells into NPC cancer cells. ²⁴ HIF‐1α, EGFR, Ki‐67, PTEN, VEGF, and EBV are considered important prognostic factors of NPC through their regulation of the hypoxic tumor microenvironment. The combination of MR images and gene expression information may be helpful for a better PFS prediction, but it has not been applied in NPC clinical practice.

To combine medical image and gene information in PFS prediction, the texture analysis method has been widely used to build nomograms, which can predict PFS and other important patient characteristics, such as tumor type and treatment response. ²⁵ , ²⁶ , ²⁷ , ²⁸ Contextual properties of the lesion region of interest (ROI) have been extracted and fitted to the logistic regression model (for a categorical amount) or the hazard regression model (for PFS). ²⁹ , ³⁰ This texture analysis method has been used to predict the PFS of nasopharyngeal carcinoma patients. ³¹ , ³² Contrast‐enhanced T1‐weighted images and T2‐weighted images are the most frequently used input for feature extraction. ³³ , ³⁴

Texture analysis‐based MR image processing enables the easy combination of information from different modalities, and the model is straightforward to explain and extend. However, it has several limitations. The performance of the prediction depends on the kernels applied to the image. A limited number of kernels may not cover all the features of the tumor, while the optimal kernel library remains unknown and may depend on the specific task. Moreover, texture analysis can only reflect characteristics inside certain ROIs, ignoring information outside the ROI and connections between different ROIs. These limitations prevent the texture analysis method from wider application in medical image processing practice.

Recently, deep learning methods have been widely used for medical image processing. ³⁵ Based on the concatenation of multiple convolution layers, activation layers and sampling layers, deep learning methods can extract features at multiple scales without needing an ROI or a feature library. ³⁶ Furthermore, deep learning methods can provide a direct prediction of PFS without the need to rely on certain survival models. Therefore, they are promising for application to PFS prediction tasks. Although neural networks have been used to integrate features into one PFS prediction, deep learning‐based MR image processing has not been applied to the PFS prediction of nasopharyngeal carcinoma patients. ³⁷ , ³⁸ In this study, we propose a PFS prediction model using residual convolutional neural network combining MR images (T1‐weighted MRI [T1w], DCE, and proton density [PD]), tumor stage, and gene expression information (HIF‐1α, EGFR, PTEN, Ki‐67, VEGF, and EB virus) to predict the PFS of nonmetathestic NPC patients. We report here the PFS prediction accuracy of this deep learning method and the comparison with the texture analysis method.

2. MATERIALS AND METHODS

2.1. Patient information

This retrospective study was approved by the Hainan General Hospital (The Affiliated Hainan Hospital of Hainan Medical University) Ethics Committee and Institutional Review Board. One hundred and fifty‐one patients diagnosed with NPC between October 2018 and October 2020 were included in the study. T1, DCE, and PD image and gene expression information for all subjects were acquired before IMRT treatment. A follow‐up study of all patients was performed after treatment to acquire PFS. Details of patient enrollment criteria, MR protocols, histopathological protocols, and IMRT treatment are described in Appendix S1, and the patient information is shown in Table 1.

TABLE 1.

Patient characteristics in training and testing dataset

Characteristics	Training dataset	Testing dataset
Data size	75	76
Age (year)	52.0 ± 11.2	49.1 ± 11.4
Sex
Male	56	58
Female	19	18
T classification
T1	0	0
T2	8	3
T3	33	42
T4	34	31
N classification
N1	24	26
N2	40	31
N3	11	19
M classification
0	70	72
1	5	4
HIF‐1 $\alpha$
≤20%	50	59
>20%	25	17
EGFR
0–1	40	38
2–3	35	38
Ki‐67
≤50%	36	32
>50%	39	44
PTEN
0	4	3
1	71	73
VEGF
0	24	33
1	51	43
EB virus
0	26	22
1	49	54
PFS (month)	11.82 ± 6.83	12.28 ± 6.64

Abbreviations: EB = Epstein‐Barr; EGFR = epidermal growth factor receptor; HIF‐1α = hypoxia inducible factor‐1‐alpha; PFS = progression free survival; PTEN = Phosphatase and tensin homolog; VEGF = Vascular Endothelial Growth Factor.

2.2. Image analysis

All image series were co‐registered to PD images, which have the highest plane resolution (0.46*0.46 mm) based on the landmark. The 3D segmentation of the primary lesion was traced manually on PD images using ITK‐SNAP (version 3.4.0, USA, http://www.itksnap.org). ⁵³ This process was performed by a clinical radiologist (R.L.) with 2 years of experience and verified by a senior radiologist (W.H.) with 11 years of experience. The radiologists were blind to the clinical and histopathological information. Representative T1w, PD, and pre‐ and post‐contrast DCE images are presented in Figure 2. All subjects were randomly divided into two groups: a training group (N = 75) and a testing group (N = 76). Two image processing methods, deep learning and texture analysis, were then performed to predict PFS. For deep learning PFS prediction, we implemented an 18‐layer 2D residual network with images of three MR modalities concatenated in channel dimension (Figure 1B). ³⁶ A linear regression model was used to combine the deep neural network output of each slice into one single output. For texture analysis PFS prediction, 120 texture features were extracted from each 3D MRI volume using the Pyradiomics toolbox. Features were selected using least absolute shrinkage and selection operator (LASSO) regression. ⁵⁴ Selected features were combined using a multivariate hazard regression model. Details of the deep neural network and texture analysis model are described in Appendix S1.

FIGURE 2.

Magnetic resonance (MR) images of a 49‐year‐old IV‐A stage (T stage:4, N stage:3, M stage:0) nasopharyngeal carcinoma (NPC) patient. (A) Pre‐contrast dynamic contrast enhancement (DCE) MR image, (B) post‐contrast (80 s after injection) DCE MR image, (C) T1‐weighted image, (D) proton density image, (E) pathology of HIF‐1α expression, and (F) pathology of epidermal growth factor receptor (EGFR) expression.

FIGURE 1.

(A) Data processing pipeline. Progression‐free survival (PFS) prediction models from magnetic resonance (MR) images were built separately from a deep learning method and texture analysis method. The prediction model further combined patient characteristics using a linear regression model to improve the performance. (B) 18‐layer residual network architecture implemented in this study.

2.3. Combination of image analysis output and patient characteristics

To combine patient characteristics with risk prediction from the texture analysis method or PFS prediction from the deep learning method, a linear regression model was trained on the same training dataset with real PFS as the target and tested on the test dataset. Two regression models were constructed: one with texture analysis output and another with deep learning output. Nine patient characteristics were considered in this study: HIF‐1α expression level, EGFR expression level, Ki‐67 expression level, PTEN expression level, VEGF expression level, EB virus infection level, patient age, gender, and tumor stage. Those patient characteristics were selected step wisely and included in the two regression models separately by minimizing Akaike's information in the likelihood ratio test.

2.4. Statistical analysis

The performance of the following five models was evaluated: deep learning prediction from MR images, texture analysis prediction from MR images, patient characteristics only, deep learning prediction with patient characteristics, and texture analysis prediction with patient characteristics. The Spearman correlation test was performed to test the correlation between real PFS and PFS estimated from these models. The concordance index (Harrell's C‐index) was calculated to test if the proposed methods can provide accurate relative PFS. ⁵⁵ The receiver operative curve (ROC) was plotted for 1‐year PFS, and the area under the curve (AUC) value was calculated to evaluate the discriminative ability of the image processing methods. The optimal cutoff was chosen to maximize sensitivity plus specificity. Furthermore, we classified the test dataset into “high‐risk” and “low‐risk” groups based on each image processing method. Finally, a log‐rank test was performed to determine the cutoff value and compare the survival curve of these two groups. All statistical analysis was performed using R statistical software (Foundation for Statistical Computing, Vienna, Austria), and a p‐value smaller than 0.05 was considered significant.

3. RESULTS

Nineteen features extracted from the texture analysis were selected for the multivariate hazard regression model. The correlations between these feature values and PFS are listed in Table 2. In the test cohort, between real PFS and predicted PFS, Harrell's C‐index of the regression model is 0.66 (95% confidence interval [CI] 0.63–0.69). For 1‐year PFS classification, the area under the curve [AUC] for the receiver operative curve [ROC] curve is 0.72 (95% CI 0.63–0.85, sensitivity = 0.85, specificity = 0.62). The log‐rank test shows there is a significant difference between high‐risk group and low‐risk groups (hazard ratio [HR] = 2.69, 95% CI 1.63–4.43, p < 0.001, z = 3.76).

TABLE 2.

Features selected in texture analysis

Name	Correlation with PFS
NonUniformity_PrecontrastDCE	0.2582
SmallAreaEmphasis_PrecontrastDCE	0.2626
Correlation_PostcontrastDCE (10s)	0.2600
Minimum_PostcontrastDCE (15s)	0.2412
90Percentile_PostcontrastDCE (20s)	0.2286
DifferenceVariance_PostcontrastDCE (20s)	0.2276
LowGrayLevelZoneEmphasis_PostcontrastDCE (20s)	0.2309
Autocorrelation_PostcontrastDCE (25s)	0.2542
InverseVariance_PostcontrastDCE (25s)	0.2295
HighGrayLevelEmphasis_PostcontrastDCE (25s)	0.2628
SmallDependenceEmphasis_PostcontrastDCE (25s)	0.2729
ShortRunHighGrayLevelEmphasis_PostcontrastDCE (25s)	0.2686
LowGrayLevelZoneEmphasis_ PostcontrastDCE (215s)	0.2469
ClusterShade_PostcontrastDCE (220s)	0.237
ClusterShade_PostcontrastDCE (220s)	0.2312
90Percentile_T1Map	0.3286
Median_T1Map	0.3234
90Percentile_ProtonDensityMap	0.2378
Uniformity_ProtonDensityMap	0.2807

Abbreviations: DCE = dynamic contrast enhanced.

The deep learning method performs better in PFS prediction than the texture analysis method. In the testing cohort between real PFS and predicted PFS, the coefficient of determination ($R^2$) is 0.44 (p < 0.001, F = 58.84, shown in Figure 3) and Harrell's C‐index is 0.80 (95% CI 0.76–0.84). For the 1‐year PFS classification, the AUC for the ROC curve is 0.85 (95% CI 0.74–0.92, sensitivity = 0.82, specificity = 0.86). The log‐rank test shows there is a significant difference between the high‐risk group and the low‐risk group (HR = 7.16, 95% CI 3.88–13.22, p < 0.001, z = 6.15).

FIGURE 3.

Correlation plot of patient progression‐free survival (PFS) and predicted PFS using (A) deep learning of magnetic resonance (MR) images and (B) deep learning of MR images + patient characteristics.

Among the nine patient characteristics, HIF‐1α expression is negatively related with PFS (R ² = 0.03, p = 0.03, F = 4.82); Ki‐67 expression is negatively related with PFS (R ² = 0.06, p = 0.004, F = 8.67); EB virous infection is negatively related with PFS (R ² = 0.08, p = 0.003, F = 14.04). EGFR expression, PTEN expression, VEGF expression, clinical stage, patient age, and gender are not significantly related to PFS (p = 0.41, 0.16, 0.38, 0.79, 0.26, and 0.89, accordingly). Between the low PFS (PFS < 12 months) and high PFS (PFS ≥ 12 months) groups, the independent‐sample t‐test shows that there is a significant difference for HIF‐1α expression (low PFS: 1.04 ± 0.94, high PFS: 0.67 ± 0.65, p = 0.0079), Ki‐67 expression (low PFS: 60.85% ± 21.65%, high PFS: 51.08% ± 19.30%, p = 0.0037), and EB expression (low PFS: 0.84 ± 0.36, high PFS: 0.52 ± 0.50, p < 0.001). There is no significant difference for EGFR expression (low PFS: 2.30 ± 0.78, high PFS: 2.28 ± 0.68, p = 0.8169), PTEN expression (low PFS: 0.97 ± 0.16, high PFS: 0.94 ± 0.24, p = 0.30), VEGF expression (low PFS: 0.64 ± 0.63, high PFS: 0.84 ± 0.82, p = 0.11), clinical stage (p = 0.46), age (low PFS: 51.78 ± 12.07 years, high PFS: 48.85 ± 10.44 years, p = 0.11), and gender (p = 0.97).

For texture analysis PFS prediction, a linear regression model combining risk prediction and HIF‐1α expression provides the minimum Akaike information (10.43). Compared with the risk prediction using MR images only, in the test cohort, for 1‐year PFS classification, the AUC of the ROC curve is increased from 0.72 to 0.76 (95% CI 0.63–0.86, sensitivity = 0.76, specificity = 0.71). Harrell's C‐index of the regression model improves from 0.66 to 0.68 (95% CI 0.64–0.70). The log‐rank test shows there is a significant difference between the high‐risk group and the low‐risk group (HR = 2.85, 95% CI 1.72–4.73, p < 0.001, z = 3.95).

For deep learning PFS prediction, a linear regression model combining a neural network output, HIF‐1α expression, and EB infection provides the minimum Akaike information (11.58). Compared with the risk prediction using MR images only, in the test cohort, for 1‐year PFS classification, the AUC of the ROC curve is increased from 0.85 to 0.88 (95% CI 0.78–0.95, sensitivity = 0.80, specificity = 0.90). Harrell's C‐index of the regression model improves from 0.80 to 0.82 (95% CI 0.78–0.85). The log‐rank test shows there is a significant difference between the high‐risk group and the low‐risk group (HR = 8.45, 95% CI 4.52–15.82, p < 0.001, z = 6.51).

As a comparison, for the linear regression model using only patient characteristics, a model combining HIF‐1α expression, Ki‐67 expression, PETN expression, and EB virus infection provides the minimum Akaike information (14.13). In the testing cohort between real PFS and predicted PFS, Harrell's C‐index is 0.63 (95% CI 0.59–0.66). For 1‐year PFS classification, the AUC for the ROC curve is 0.71 (95% CI 0.56–0.81, sensitivity = 0.71, specificity = 0.66). The log‐rank test shows there is a significant difference between the high‐risk group and the low risk group (HR = 2.37, 95% CI 1.40–4.00, p < 0.001, z = 3.09).

Statistical analysis of the five models are summarized in Tables 3, 4, 5, 6: deep learning prediction from MR images, texture analysis prediction from MR images, patient characteristics only, deep learning prediction with patient characteristics, and texture analysis prediction with patient characteristics. Deep learning prediction with patient characteristics provides the highest Harrell's C‐index (0.82, 95% CI 0.78–0.85). For 1‐year PFS classification, deep learning prediction with patient characteristics provides the highest AUC (0.88), which is significantly higher than for texture analysis prediction with patient characteristics (AUC = 0.76, p = 0.007), texture analysis prediction without patient characteristics (AUC = 0.72, p < 0.001), and patient characteristics only (AUC = 0.71, p < 0.001) but not significantly higher than for deep learning prediction (AUC = 0.85, p = 0.24). The ROC curves for the two methods are presented in Figure 4. The log‐rank test shows that deep learning prediction with patient characteristics has the largest difference between low‐risk and high‐risk groups (HR = 24.79). Survival possibility plots with time for these methods are shown in Figure 5.

TABLE 3.

C index for PFS prediction in the training dataset

	Value	95% CI
DL	0.85	0.80–0.89
TA	0.69	0.66–0.74
DL (with patient characteristics)	0.88	0.84–0.93
TA (with patient characteristics)	0.72	0.68–0.76
Patient characteristics only	0.65	0.61–0.68

Abbreviations: DL = deep learning method; TA = texture analysis method.

TABLE 4.

C index for PFS prediction in test dataset

	Value	95% CI
DL	0.80	0.76–0.84
TA	0.66	0.63–0.69
DL (with patient characteristics)	0.82	0.78–0.85
TA (with patient characteristics)	0.68	0.64–0.70
Patient characteristics only	0.62	0.60–0.65

Abbreviations: DL = deep learning method; TA = texture analysis method.

TABLE 5.

Spearman correlation with PFS value in the test dataset

	R 2	p‐value	F‐value
DL	0.44	<0.001	58.84
TA	0.07	0.03	5.06
DL (with patient characteristics)	0.53	<0.001	82.14
TA (with patient characteristics)	0.14	<0.001	12.14
Patient characteristics only	0.02	0.23	1.49

Note: For DL and TA, degree of freedom = 75. For DL with patient characteristics, degree of freedom = 73; for TA with characteristics, degree of freedom = 74; for patient characteristics only, degree of freedom = 72.

Abbreviations: DL = deep learning method; TA = texture analysis method.

TABLE 6.

ROC analysis for 1‐year PFS prediction in the test dataset

	AUC	Sensitivity	Specificity
DL	0.85 (0.74–0.92)	0.82 (0.67–0.93)	0.86 (0.72–0.94)
TA	0.72 (0.63–0.85)	0.85 (0.70–0.94)	0.62 (0.46–0.76)
DL (with patient characteristics)	0.88 (0.78–0.95)	0.80 (0.62–0.91)	0.90 (0.78–0.97)
TA (with patient characteristics)	0.76 (0.63–0.86)	0.76 (0.60–0.88)	0.71 (0.56–0.83)
Patient characteristics only	0.71 (0.57–0.82)	0.71 (0.52–0.84)	0.66 (0.52–0.80)

Note: 95% confidence interval in parentheses.

Abbreviations: DL = deep learning method. TA = texture analysis method.

FIGURE 4.

One‐year progression‐free survival (PFS) classification receiver operative curve (ROC) plot of deep learning prediction from magnetic resonance (MR) images, texture analysis prediction from MR images, deep learning prediction with patient characteristics, texture analysis prediction with patient characteristics, and patient characteristics only.

FIGURE 5.

Kaplan–Meier survival plot of (A) deep learning prediction from magnetic resonance (MR) images, (B) texture analysis prediction from MR images, (C) deep learning prediction of MR images + patient characteristics, (D) texture analysis prediction of MR images + patient characteristics, and (E) linear regression prediction using patient characteristics only.

4. DISCUSSION

With the rapid development of computer and graphics processing units, various advanced image processing methods, such as texture analysis methods and deep learning methods, have been applied to medical imaging processing to solve clinical problems. In this study, we report the performance of deep learning methods in nasopharyngeal carcinoma PFS time estimation and the comparison with the traditional regression model based on features extracted from texture analysis. The results show that deep learning methods can provide a more accurate PFS prediction, and the performance can be further improved by combining patient gene expression information.

Currently, nasopharyngeal carcinoma PFS prediction from MR images is mainly based on texture analysis. ³¹ Among the texture analysis‐based PFS prediction models, two regression models using features from contrast‐enhanced T1 images achieved C‐indexes of 0.72 and 0.75 ³⁹ , ⁴⁰ ; two regression models combining features from contrast‐enhanced T1‐weighted images and T2‐weighted images achieved C‐indexes of 0.73 and 0.79 ³⁹ , ⁴¹ ; one regression model combining features from T1‐weighted images, contrast‐enhanced T1‐weighted images, and T2‐weighted images achieved a C‐index of 0.74. ⁴² Compared to the texture analysis method, deep learning‐based MR image processing does not rely on a kernel library or manually drawn ROI. Therefore, it may capture more features and is fully automatic. In this study, significant improvement in NPC prediction was achieved by replacing texture analysis with deep learning.

Three MR modalities are involved in this study: T1w, PD, and DCE MRI. All three types of images contribute features to the hazard regression model. High‐resolution T1w and PD images can provide tumor morphology information, while DCE images can provide vessel permeability and flow information. Adding different MR modalities into the model may further improve the PFS prediction accuracy. Diffusion‐weighted imaging (DWI) images have been shown to be useful for understanding NPC treatment response, and T2‐weighted images can be used to predict locoregional recurrence and distant metastasis. ⁴³ , ⁴⁴ Moreover, quantitative parameters derived from post‐processing of MR images, such as perfusion parameters derived from DCE‐MRI or arterial spin labeling images and susceptibility derived from multi‐echo gradient reversal images, may also reflect tumor properties and help predict PFS. ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ Features from these modalities are promising coupled with the PFS prediction model.

Nine patient characteristics are included in our study: expression level of five genes (HIF‐1α, EGFR, Ki‐67, PTEN, and VEGF), EB infection level, patient age, gender, and tumor stage. HIF‐1α expression, Ki‐67 expression, and EB virus infection level show significant relationships with PFS. Adding these patient characteristics into a PFS prediction model improves the prediction accuracy, but the improvement is not significant. One possible reason is that these patient characteristics may affect the morphology and perfusion of tumors and further affect the MR image. Therefore, the MR images are not independent of patient characteristics. Correlation analysis of MR images and patient characteristics should be studied in a larger clinical cohort to better understand the value of these patient characteristics for PFS prediction.

There are several limitations of this study. First, constrained by data size, the enrolled patients were not separated into subgroups based on tumor stage or type, which may influence the correlation of MR images and patient characteristics with PFS. Constructing a tumor stage or type‐specific PFS prediction model, especially a model for advanced PFS, may contribute to the design of treatment and follow‐up of the patients. Second, motion artifacts may exist between different MR modalities and timeframes in DCE MRI scans, affecting the performance of deep learning and the texture analysis image processing pipeline. A registration algorithm for nasopharyngeal region MR should be developed and applied to minimize the motion artifact. Third, for DCE MR images, the slice thickness is large, and the temporal resolution is low, which may cause the loss of features. Optimized sampling patterns and compressed sensing techniques may accelerate the acquisition and eventually help with the feature extraction from DCE MRI. ⁵² Fourth, limited by GPU memory, we implemented an 18‐layer 2D residual network. Increasing the depth of the network or changing the 2D network to 3D may help to improve its performance further. Fifth, this study was performed in a single institution with only three MR modalities from the same scanner. Acquiring multicenter data from different MR systems is needed to further validate our algorithm. Sixth, only 1‐year PFS was acquired in this study. Understanding long‐term survival rates and overall recurrence‐free survival rates is important for patient care. Follow‐up of these patients is needed to examine whether the proposed algorithm can be applied to 3‐ or 5‐year PFS and overall recurrence‐free survival rate prediction. These issues need to be addressed in further studies.

In conclusion, deep learning methods are promising for PFS prediction from MR images and patient characteristics. Future work may include exploring the PFS prediction ability of different MR modalities, such as T2‐weighted images, and improving the PFS prediction accuracy in specific tumor stages, especially TNM III and IV stages.

ACKNOWLEDGMENTS

We would like to thank all the patients and the personnel involved in this study.

AUTHOR CONTRIBUTION

Qihao Zhang: experiment design; data processing; manuscript drafting. Gang Wu: data collection; manuscript review. Qianyu Yang: experiment design; manuscript review. Ganmian Dai: data processing; manuscript review. Tiansheng Li: data processing; manuscript review. Pianpian Chen: data collection; manuscript review. Jiao Li: data collection; manuscript review. Weiyuan Huang: data collection; experiment design; manuscript drafting.

CONFLICT OF INTEREST

The authors have nothing to disclose.

ETHICS STATEMENT

This retrospective study was approved by the Hainan General Hospital (The Affiliated Hainan Hospital of Hainan Medical University) Ethics Committee and Institutional Review Board.

Funding information: This study is supported by Hainan Province Clinical Medical Center.

Informed Consent: N/A.

Registry and the Registration No. of the study/trial: N/A.

Animal Studies: N/A.

Supporting information

Appendix S1.

Zhang Q, Wu G, Yang Q, et al. Survival rate prediction of nasopharyngeal carcinoma patients based on MRI and gene expression using a deep neural network. Cancer Sci. 2023;114:1596‐1605. doi: 10.1111/cas.15704