A deep learning-based radiomic nomogram for prognosis and treatment decision in advanced nasopharyngeal carcinoma: A multicentre study

Abstract

Background

Induction chemotherapy (ICT) plus concurrent chemoradiotherapy (CCRT) and CCRT alone were the optional treatment regimens in locoregionally advanced nasopharyngeal carcinoma (NPC) patients. Currently, the choice of them remains equivocal in clinical practice. We aimed to develop a deep learning-based model for treatment decision in NPC.

Methods

A total of 1872 patients with stage T3N1M0 NPC were enrolled from four Chinese centres and received either ICT+CCRT or CCRT. A nomogram was constructed for predicting the prognosis of patients with different treatment regimens using multi-task deep learning radiomics and pre-treatment MR images, based on which an optimal treatment regimen was recommended. Model performance was assessed by the concordance index (C-index) and the Kaplan-Meier estimator.

Findings

The nomogram showed excellent prognostic ability for disease-free survival in both the CCRT (C-index range: 0.888-0.921) and ICT+CCRT (C-index range: 0.784-0.830) groups. According to the prognostic difference between treatments using the nomogram, patients were divided into the ICT-preferred and CCRT-preferred groups. In the ICT-preferred group, patients receiving ICT+CCRT exhibited prolonged survival over those receiving CCRT in the internal and external test cohorts (hazard ratio [HR]: 0.17, p<0.001 and 0.24, p=0.02); while the trend was opposite in the CCRT-preferred group (HR: 6.24, p<0.001 and 12.08, p<0.001). Similar results for treatment decision using the nomogram were obtained in different subgroups stratified by clinical factors and MR acquisition parameters.

Interpretation

Our nomogram could predict the prognosis of T3N1M0 NPC patients with different treatment regimens and accordingly recommend an optimal treatment regimen, which may serve as a potential tool for promoting personalized treatment of NPC.

Funding

National Key R&D Program of China, National Natural Science Foundation of China, Beijing Natural Science Foundation, Strategic Priority Research Program of CAS, Project of High-Level Talents Team Introduction in Zhuhai City, Beijing Natural Science Foundation, Beijing Nova Program, Youth Innovation Promotion Association CAS.

Research in Context.

Evidence before this study

We searched PubMed, Google Scholar and Web of Science for papers published from database inception to May, 2021, with the terms (“radiomics” OR “deep learning” OR “machine learning” OR “MR-radiomics” OR “radiomic nomogram” OR “imaging predictor”) AND (“NPC” OR “nasopharyngeal carcinoma” OR “prognosis and treatment decision” OR “prognosis and treatment response” OR “treatment decision” OR “treatment response” OR “benefit” OR “prognosis”), with no language restrictions and source restrictions. To the best of our knowledge, there are no published studies so far that have used radiomics to predict the prognosis of patients with cancers including NPC with different treatment regimens and accordingly recommend an optimal treatment regimen. Studies in treatment decision of NPC using radiomics have focused on risk stratification of a prognostic model/factor and development of predictive signatures for treatment response, but did not provide the prognosis of patients with different treatment regimen. None of the studies use multi-task deep learning radiomics.

Added value of this study

Our study is the first multi-task deep learning study that could predict prognosis and concurrently recommend treatment regimens for NPC patients, which was validated in both internal and external test cohorts from a multicentre and large-scale dataset, supporting the move toward personalized and optimized management of NPC.

Implications of all the available evidence

Our study suggests the potential of multi-task deep-learning radiomics for developing more comprehensive prognostic model with personalized treatment recommendations for NPC, which also offers a promising start for treatment decisions and prognostic prediction in other cancers. By providing the prognosis of patients with different treatment regimens and accordingly recommending an optimal treatment regimen, integrating our model into routine clinical practice would promote personalized treatment and optimize management of NPC.

Alt-text: Unlabelled box

1. Introduction

Nasopharyngeal carcinoma (NPC) is a radiosensitive epithelial malignancy associated with the risk of locoregional and distant aggressiveness, with the highest prevalence reported for the populations of East and Southeast Asia, and North Africa [1]. Of the 87,000 newly diagnosed NPC annually, over 70% of them are classified as advanced disease [2]. Patients with advanced NPC have several treatment options, including induction chemotherapy (ICT) plus concurrent chemoradiotherapy (CCRT) and CCRT alone [3,4]. Currently, the choice of individualized treatment regimen remains equivocal in clinical practice [3,5].

In recent decades, some prognostic biomarkers, such as mRNA [6], plasm protein [7], and tumour-infiltrating lymphocytes [8], had been developed for advanced NPC. However, these prognostic signature are not predictive biomarkers for treatment response, as they did not inform therapy strategy before treatment [9,10]. Therefore, it is very necessary to examine the impact of treatment regimen on prognosis of NPC patients to help optimize treatment decision.

Recently, the use of radiomics in oncology, which allows in-depth characterization of tumour phenotypes using a computerized imaging analysis algorithm, has garnered increasing attention [11,12]. The introduce of deep learning enables radiomics to perform end-to-end modelling of medical data and multi-task learning for multiple clinical tasks, which demonstrated excellent results [13], [14], [15]. Emerging evidence has confirmed that radiomics had potential value in predicting prognosis and treatment response in NPC [16], [17], [18]. However, no previous study has used multi-task learning to simultaneously predict prognosis and make treatment recommendations for NPC patients.

The aim of this study was to use multi-task deep-learning radiomics to develop simultaneously prognostic and predictive signatures from pre-treatment magnetic resonance (MR) images of NPC patients, and to construct a combined prognosis and treatment decision nomogram (CPTDN) for predicting the prognosis of NPC patients with different treatment regimens and accordingly recommending an optimal treatment regimen. In this study, we focused on stage T3N1M0 NPC patients, the largest subgroup of advanced NPC, and collected a large multicentre dataset used for model development and validation.

2. Methods

2.1. Patients

The ethical review board of Institute of Automation, Chinese Academy of Sciences approved this retrospective study (ID: IA-202043) and waived the requirement of informed consent.

The overall study design is shown in Fig. 1. We retrospectively reviewed the charts of patients with NPC diagnosed at the four Chinese centres (centre 1: Sun Yat-sen University Cancer Centre, centre 2: First People's Hospital of Foshan Affiliated to Sun Yat-sen University, centre 3: Fifth Affiliated Hospital of Sun Yat-sen University, centre 4: Guilin Medical University Affiliated Hospital) between January 2010 and June 2017. A total of 1872 candidate eligible patients with stage T3N1M0 NPC and available diagnostic MR images were included, according to the inclusion and exclusion criteria.

Fig. 1.

Deep-learning radiomics workflow. (a) Extracting deep radiomic features from multi-sequence magnetic resonance images. (b) Building prognostic and predictive signatures using a unified fully connected neural network. (c) Constructing and validating a prognostic nomogram that could predict the prognosis of patients with different treatment regimens. (d) Supporting treatment decision using the prognostic difference between treatments based on the nomogram. DFS, disease-free survival.

The inclusion criteria: had diagnosed T3N1M0; treated by either ICT+CCRT or CCRT alone; received intensity-modulated radiotherapy; underwent MRI scans within 2 weeks before treatment.

The exclusion criteria: received adjuvant chemotherapy, targeted therapy, or biotherapy during the course of radical treatment; with previous chemotherapy or radiotherapy or other malignant tumours; had artifacts, blurs, faults, and disordered slices in the MR image.

Data on demographic characteristics, pre-treatment plasma Epstein-Barr virus DNA (pre-EBV DNA) levels, and treatment regimen were gathered from the medical records of each centre. All patients were treated by either CCRT or ICT+CCRT. To reduce the impact of clinicians’ judgment on the choice of treatment regimen, a nonparametric matching method was used to screen out patients with similar baseline characteristics in the CCRT and ICT+CCRT groups at a ratio of 1:1 [19] (Appendix A). Matched patients from centre 1 were divided into a training cohort and an internal test cohort, and those from the other three centres were divided into an external test cohort.

The primal endpoint is disease-free survival (DFS), the time from the date of initial diagnose until either the date of disease progression or death from any cause. The protocol of disease progression identification, including local and regional recurrence or distant metastasis, was the same as previously reported [17]. Tumour staging was determined based on the American Joint Committee on Cancer TNM Staging Manual, eighth Edition [20]. Treatment details are summarised in Appendix A.

2.2. MR acquisition and tumour segmentation

In two weeks prior to any anti-tumour therapy, all patients underwent three sequences of MRI (T1-weighted, T2-weighted, and contrast enhanced T1-weighted) scanning. The MR acquisition protocols (including the details of manufacturer and magnetic field strength, among others) are shown in Appendix B. For each MR sequence, the primary tumour was manually contoured on each axial MR slice by two experienced radiologists (X.L.F. and F.Z.).

2.3. Feature extraction and radiomic signature building

Considering scanner-dependent bias in image intensity, the MR images were interpolated and normalized to match the reference template [21]. We used the SE-ResNet architecture [22] as the backbone network combined with the multiple instance learning method [23] to extract the deep radiomic features from the MR images. Afterward, taking the deep radiomic features from three MR sequences as input, a unified fully connected neural network, which consists of a shared backbone network and two task-specific subnetworks, was constructed to simultaneously predict prognosis and treatment response. The output of the subnetwork for predicting prognosis (Prognostic-score) was the estimate of the risk of disease progression, whereas that of another subnetwork (Predictive-score) was the estimate of the relative risk of disease progression associated with receiving ICT+CCRT vs. receiving CCRT. In the subnetwork that predicted treatment response, interactions between treatments and activation units were considered, using the modified covariate method [18]. Cox partial log likelihood [24] was chosen as the loss function to train feature extraction networks and signature building network in the training cohort. Specific architectures and training details of these networks were shown in Appendix C.

2.4. CPTDN construction

Univariate Cox proportional hazards regression analysis was used to evaluate the association between clinical characteristics and DFS and response to treatment regimen in the training cohort. Incorporating independent clinical factors, Prognostic-score, and Predictive-score, a CPTDN was developed using multivariate Cox proportional hazards regression (MCox) method and was plotted for the convenient of clinicians. Using the CPTDN, the prognosis predictions of each patient were calculated when it receives different treatment regimens, based on which the prognostic difference was used to make a personalized treatment recommendation.

To compare to CPTDN in prognostic prediction, three clinical models were built, using MCox method in all patients ($Mode{l}_{all}^{clin}$), those receiving CCRT ($Mode{l}_{ccrt}^{clin}$) and those receiving ICT+CCRT ($Mode{l}_{ict+ccrt}^{clin})$ of the training cohort. In addition, $Mode{l}_{all}^{clin+prog}$ was built using MCox method in the training cohort to assess the incremental prognostic value of prognostic-score in clinical practice.

2.5. Performance assessment

Harrell's concordance index (C-index) and hazard ratio (HR) estimates were used to assess and test prognostic performance of the proposed models in all cohorts. The quality of Predictive-score was assessed by its $p_{interaction}$ value, indicative of treatment interaction.

Furthermore, the calibration curve was plotted to evaluate the agreement between the observed survival and the CPTDN-predicted survival. Subgroup analysis using clinical factors and MR acquisition parameters was performed to assess the stability of CPTDN. In addition, we evaluated the prognostic performance of CPTDN in terms of second clinical endpoints, including overall survival, distant metastasis-free survival, and locoregional relapse-free survival.

2.6. Statistical analysis

The survival curves were drawn using Kaplan-Meier method. The p-value of HR estimate was computed using the log-rank test. In the MCox analysis, a hybrid mode of forward and backward stepwise selection was applied to select independent features using Akaike's information. A two-sided p<0.05 means that the corresponding estimate reaches significant difference.

Open-source R software v3.6.1 was used to conduct statistical analysis, and open-source Python v3.6.8 and TensorFlow v1.12 was used to implement deep learning network models. A detailed description of the R packages used is shown in Appendix C.

2.7. Role of the funding source

The funders had no role in the study design, data collection, analysis, patient recruitment, writing of the manuscript, the decision to submit the manuscript for publication, or any aspect pertinent to the study. All authors had full access to the full data in the study and accept responsibility to submit for publication.

3. Results

3.1. Clinical characteristics

According to sex, age, smoking status, drinking status, family history of cancer, haemoglobin levels, tumour volume, and pre-EBV DNA levels, 1008 matched NPC patients with stage T3N1M0 in centre 1 were selected and divided randomly into the training (n=684) and internal test (n=324) cohorts. The external test cohort comprised of 198 matched target patients from three other centres. All matched patients had balanced baseline clinical characteristics between treatments (Table A1). Baseline clinical characteristics are summarized in Table 1; they did not show significant differences between the cohorts, except for the pre-EBV DNA levels. Median follow-up was 64.0 months (interquartile range [IQR]: 53.4-78.3), 65.7 months (IQR: 52.9-77.6), and 63.3 months (IQR: 50.9-76.6) in the training, internal test, and external test cohorts, respectively. At the last follow-up, 107/684 (15.6%), 55/324 (17.0%), and 35/198 (17.7%) patients had experienced a confirmed disease progression in each of the cohort, respectively.

Table 1.

Baseline characteristics in the training, internal validation, and external validation cohorts.

Characteristics	Training (N = 684)	Internal test (N = 324)	External test (N = 198)	p-value
Age (years), median (range)	42 (14-72)	41 (15-65)	43 (13-72)	0.33
Sex, No. (%)				0.63
Male	464 (67.8)	229 (70.7)	134 (67.7)
Female	220 (32.2)	95 (29.3)	64 (32.3)
Haemoglobin1 (g/L), No. (%)
Normal	599 (87.6)	283 (87.3)	152 (86.4)	0.91
Abnormal	85 (12.4)	41 (12.7)	24 (13.6)
Unknown	0	0	22
Family history of cancer, No. (%)				0.72
Yes	189 (27.6)	82 (25.3)	44 (27.7)
No	495 (72.4)	242 (74.7)	115 (72.3)
Unknown	0	0	39
Smoking, No. (%)				0.73
Yes	205 (30.0)	105 (32.4)	50 (31.4)
No	479 (70.0)	219 (67.6)	109 (68.6)
Unknown	0	0	39
Drinking, No. (%)				0.33
Yes	72 (10.5)	40 (12.3)	23 (14.5)
No	612 (89.5)	284 (87.7)	136 (85.5)
Unknown	0	0	39
WHO pathology type, No. (%)				0.54
I-II	7 (1.0)	2 (0.6)	3 (1.6)
III	677 (99.0)	322 (99.4)	179 (98.4)
Unknown	0	0	16
Albumin (g/L), No. (%)				0.88
Normal (40-55)	630 (92.1)	301 (92.9)	146 (91.8)
Abnormal (other)	54 (7.9)	23 (7.1)	13 (8.2)
Unknown	0	0	39
C-reaction protein (mg/L), No. (%)				0.97
Normal (≤3)	505 (73.8)	238 (73.5)	116 (73.0)
Abnormal (other)	179 (26.2)	86 (26.5)	43 (27.0)
Unknown	0	0	39
Lactate dehydrogenase (U/L), No. (%)				0.52
Normal (120-250)	634 (92.7)	304 (93.8)	152 (91.0)
Abnormal (other)	50 (7.3)	20 (6.2)	15 (9.0)
Unknown	0	0	31
Treatment				1.00
CCRT	341 (49.9)	163 (50.3)	99 (50.0)
ICT+CCRT	343 (50.1)	161 (49.7)	99 (50.0)
pre-EBV DNA (copy/mL), No. (%)				0.0087
< 4000	427 (62.4)	212 (65.4)	91 (51.7)
≥ 4000	257 (37.6)	112 (34.6)	85 (48.3)
Unknown	0	0	22
Tumour volume (mL), mean	29.2	28.6	31.4	0.88

¹For males, normal haemoglobin levels are in the range of 130-175 g/L; the equivalent values for females are in the range of 115-150 g/L.

WHO, World Health Organization; pre-EBV DNA, pre-treatment plasma Epstein-Barr virus DNA.

3.2. Development and validation of radiomic signatures

Prognostic-score and Predictive-score were developed simultaneously by the unified fully connected neural network. Prognostic-score exhibited a good predictive value for DFS in the training cohort (C-index: 0.772, 95% confidence interval [CI]: 0.721-0.822). Similar prognostic ability of Prognostic-score was obtained in the internal test (C-index: 0.733, 95% CI: 0.657-0.809), and external test (C-index: 0.681, 95% CI: 0.568-0.793) cohorts. Predictive-score exhibited a strong association with treatment response in all cohorts (all $p_{interaction}<0.001$ [Wald test]). Furthermore, Predictive-score did not exhibit a consistently significant association with any clinical factor in all cohorts (Table D1).

3.3. Construction and prognostic performance of CPTDN

Univariate Cox proportional hazards regression analysis (Fig. D1) demonstrated that age (HR: 1.25 [per 10 years], log-rank test: p = 0.020) and pre-EBV DNA levels (HR: 2.08, log-rank test: p < 0.001) were significantly associated with DFS, and that no clinical factor was significantly associated with treatment response (all $p_{interaction}>0.05\left[\text{Wald}\text{test}\right]).$ Incorporating Prognostic-score, interaction item of treatment with Predictive-score, pre-EBV DNA levels, and age, CPTDN was constructed for DFS prediction in the training cohort using the MCox analysis and is plotted in Fig. 2a. Meanwhile, $Mode{l}_{all}^{clin}$ was constructed using age and pre-EBV DNA levels; $Mode{l}_{ccrt}^{clin}$ was constructed using tumour volume and pre-EBV DNA levels; $Mode{l}_{ict+ccrt}^{clin}$ was constructed using age and pre-EBV DNA levels; and $Mode{l}_{all}^{clin+prog}$ was constructed using age, Prognostic-score and pre-EBV DNA levels. CPTDN obtained the best prognostic ability in the treatment-specific subgroup of all cohorts (Fig. 3). Furthermore, the C-index of CPTDN (training: 0.868 [95% CI 0.833-0.903], internal test: 0.856 [0.807-0.905], external test: 0.851 [0.787-0.914]) was also highest compared to other models in all cohorts (Table D2).

Fig. 2.

Combined prognosis and treatment decision nomogram (CPTDN) and its calibration curves. (a) A CPTDN was built for predicting disease-free survival (DFS) of patients with different treatment regimens and accordingly recommend an optimal treatment regimen. The calibration curves of the CPTDN were plotted for predicting the 3-year (b) and 5-year DFS rate (c) in three cohorts (n=1206). The CPTDN could noninvasively recommend an optimal treatment regimen and predict patient prognosis. For example, consider a 40-year-old target patient with stage T3N1M0, a pre-EBV DNA of >4000 copy/mL, a Prognostic-score of 0.0, and a Predictive-score of -0.4; If the patient received ICT+CCRT, the patient's total score would be 122 (21+15+41+45), for which the 5-year DFS rate is estimated at 81%. If the patient received CCRT, the patient's score would be 147 (21+15+41+70), for which the 5-year DFS rate is estimated at 63%. The 5-year prognostic difference is 18%, so the CPTDN would recommend ICT+CCRT and provide a 5-year DFS rate of 81%. CCRT, concurrent chemoradiotherapy; ICT, induction chemotherapy; pre-EBV DNA, pre-treatment plasma Epstein-Barr virus DNA. HL test: Hosmer-Lemeshow test.

Fig. 3.

Comparison of prognostic models’ performance on different treatment-specific subgroups. The vertical line over the bar presents the confidence interval of the C-index. CCRT, concurrent chemoradiotherapy; ICT, induction chemotherapy.

CPTDN presents good agreement with the actual observed DFS rate at 3-year and 5-year (Fig. 2b and 2c). Risk stratification analysis demonstrated that CPTDN could successfully classify patients into the low-risk and high-risk groups with significant differences in DFS (log-rank test: p <0.05, Fig. 4). Similarly, CPTDN could identify patients with good prognosis for second clinical endpoints in the combined internal and external test cohort (Fig. D2). In addition, subgroup analysis demonstrated that CPTDN retained the capacity for risk stratification for DFS (Fig. D3) when considering clinical factors and MR parameters, including sex, C-reaction protein levels (normal or abnormal), smoking status (yes or no), family history of cancer (yes or no), tumour volume ($\le$29 mL or >29 mL), magnetic field strength and manufacturer of MR scanner (<3.0T or =3.0T).

Fig. 4.

Risk stratification analysis of CPTDN for predicting disease-free survival in three cohorts (n=1206). (a) Kaplan-Meier curves of disease-free survival in the entire group, (d) patients receiving CCRT, and (g) patients receiving ICT+CCRT in the training cohort. (b) Kaplan-Meier curves of disease-free survival in the entire group, (e) patients receiving CCRT, and (h) patients receiving ICT+CCRT in the internal test cohort. (c) Kaplan-Meier curves of disease-free survival in the entire group, (f) patients receiving CCRT, and (i) patients receiving ICT+CCRT in the external test cohort. All Kaplan-Meier curves were stratified by the mean total point (100) of CPTDN outputs. CCRT, concurrent chemoradiotherapy; ICT, induction chemotherapy.

3.4. Performance for treatment decision of CPTDN

According to the CPTDN, the prognostic difference was calculated using the difference in DFS rate of receiving ICT+CCRT vs. receiving CCRT. The 5-year prognostic difference of receiving ICT+CCRT vs. receiving CCRT divided patients into the ICT-preferred group and CCRT-preferred group, using a threshold of 0% (Fig. 5). In the ICT-preferred group (difference >0%), ICT+CCRT was a preferred treatment regimen with an improved DFS compared to CCRT in the training (HR: 0.21, 95% CI: 0.11-0.38, log-rank test: p < 0.001), internal test (HR: 0.17, 95% CI: 0.07-0.42, log-rank test: p < 0.001) and external test (HR: 0.24, 95% CI: 0.07-0.90, log-rank test: p = 0.022) cohorts. While in the CCRT-preferred group (difference $\le$0%), CCRT was a preferred treatment regimen compared to ICT+CCRT in the training (HR: 5.34, 95% CI: 2.49-11.47, log-rank test: p < 0.001), internal test (HR: 6.24, 95% CI: 2.14-18.19, log-rank test: p < 0.001) and external test (HR: 12.08, 95% CI: 2.74-53.19, log-rank test: p < 0.001) cohorts. Moreover, in the treatment decision analysis of the 5-year prognostic difference with 2% and 5% as cut-off points, it remained an effective indicator for identifying patients who were suitable for CCRT or ICT+CCRT (Fig. D4 and D5). Furthermore, CPTDN could identify patients who were suitable for CCRT or ICT+CCRT for second clinical endpoints in the combined internal and external test cohort (Fig. 6). When stratified by sex, age ($\le$42 years or >42 years), tumour volume ($\le$29 mL or >29 mL), and magnetic field strength of MR scanner (<3.0T or =3.0T), the 5-year prognostic difference still identified patients who were suitable for CCRT or ICT+CCRT (Fig. D6).

Fig. 5.

Kaplan-Meier curves of disease-free survival according to dichotomized 5-year prognostic difference in all cohorts (n=1206). (a) Comparisons of the ICT+CCRT and CCRT groups in the ICT-preferred group (difference >0%), and (b) the CCRT-preferred group (difference ≤0%) in the training cohort. (c) Comparisons of the ICT+CCRT and CCRT groups in the ICT-preferred (difference >0%) group, and (d) CCRT-preferred group (difference ≤0%) in the internal test cohort. (e) Comparisons of the ICT+CCRT and CCRT groups in the ICT-preferred group (difference >0%), and (f) CCRT-preferred group (difference ≤0%) in the external test cohort. CCRT, concurrent chemoradiotherapy; ICT, induction chemotherapy.

Fig. 6.

Kaplan-Meier curves for second clinical endpoints according to dichotomized 5-year prognostic difference in the combined internal and external test cohort (n=500). (a) Comparisons of overall survival between the ICT+CCRT and CCRT groups in the ICT-preferred group (difference >0%), and (b) CCRT-preferred group (difference ≤0%). (c) Comparisons of distant metastasis-free survival between the ICT+CCRT and CCRT groups in the ICT-preferred (difference >0%) group, and (d) CCRT-preferred group (difference ≤0%). (e) Comparisons of locoregional relapse-free survival between the ICT+CCRT and CCRT groups in the ICT-preferred group (difference >0%), and (f) CCRT-preferred group (difference ≤0%). CCRT, concurrent chemoradiotherapy; ICT, induction chemotherapy.

4. Discussion

In this large multicentre study, we developed a nomogram (CPTDN) for prognosis and treatment decision in NPC patients with stage T3N1M0. The CPTDN was constructed by the predictive radiomic signature and independent prognostic factors, which could predict the prognosis of patients with different treatment regimens and accordingly recommend an optimal treatment regimen. Our CPTDN identified patients most likely to benefit from specific treatment regimens and presented better prognostic ability than the clinical models developed on different treatment-specific groups. These findings indicate that the proposed model may serve as a tool for promoting personalised treatment and optimizing the management of NPC patients. We have developed a user-friendly web browser-based tool for our CPTDN on an open-access website (http://www.radiomics.net.cn/post/135) to facilitate its validation and application.

Previous radiomics/deep learning studies in NPC separately explored the prognosis value or predictive value of clinical factors/models. Although some studies found that imaging-based prognostic signatures were predictive of treatment outcomes, they did not take treatment factor into account in the development of prognostic models [16,25]. Given the fact that treatment regimen may largely affect clinical outcomes, it is necessary to consider treatment factor when analysing prognosis. Some tailored predictive signatures could identify the patients who would benefit most from adjuvant therapy or induction therapy, but they could not give the individualized prognostic evaluation for a patient receiving the specific treatment regimen [10,18]. Therefore, these studies could not predict the prognosis of patients with different treatment regimen. In contrast, Our CPTDN incorporating prognostic and predictive factors could provide individualized evaluation of DFS in different treatment regimen and accordingly recommend optimal treatment regimens, CCRT alone or ICT+CCRT, to NPC patients. Our CPTDN remained good prognostic and predictive ability for second clinical endpoints. Moreover, the prognostic and predictive stability of our CPTDN was further validated in different subgroups stratified by clinical factors and MR acquisition parameters. To the best of our knowledge, this is the first multi-task deep learning study aimed at simultaneously predicting prognosis and recommending treatment regimens for NPC patients.

In our study, we found that Prognostic-score and Predictive-score showed a large incremental value in prognostic prediction of DFS, which validated the prognostic and predictive value of multi-task deep-learning radiomics and emphasized the necessity of considering treatment factor in prognostic prediction. Compared the single-task radiomic model (C-index: 0.759) [17], the multi-task radiomic model integrating Prognostic-score and Predictive-score obtained better prognostic results (C-index: 0.836 in the combined internal and external test cohort). To provide insights into what the multi-task deep-learning model learned, we performed gradCAM analysis [26] on patients’ MR images (Fig. D7). This analysis revealed that the primary tumour regions of the MR images are most responsible for the Prognostic-score in high-risk patients, whereas the regions adjacent to the primary tumour are most responsible for this signature in low-risk patients. As past studies indicate, both intra-tumoral and peritumoral regions offer information valuable for prognosis [27]. In the areas surrounding the tumour mass, immune signatures, such as lymphangiogenic activity and lymphocytic infiltration can be detected [28]. In high-risk patients with poor survival, the primary tumour tends to be more aggressive and the peritumoral immune response rate tends to be low. In contrast, in low-risk patients with good survival, the primary tumour is less aggressive and the peri-tumour immune response rate is high. This may explain our findings. When evaluating the Predictive-score, the multi-task deep-learning model focused on the primary tumour regions of the CCRT-preferred group and on the peritumoral regions of the ICT-preferred group. This differentiation may be due to the fact that CCRT-preferred patients tend to have vessel malfunction, which may cause hypoxia inside the tumour and drive chemoresistance [29,30]. However, in ICT-preferred patients, there are more functional vessels around the tumour that guarantee effective blood supply and drug delivery [29].

This study has three main strengths. First, this multicentre study offers a promising start for developing more comprehensive prognostic model with personalized treatment recommendations for NPC. The present study design and methods can be applied to treatment decisions and prognostic prediction in other cancer types. Second, target delineation in radiotherapy ensures the repeatability of the extracted features, as the feature extraction network uses tumour-centred patches as input, which improves the convenience and feasibility of our model in clinical practice. Finally, our study validated the incremental value of multi-task learning in prognosis prediction, paving the path for the exploration of multi-task learning in fields, such as tumour staging and disease surveillance.

Nevertheless, this study has some limitations, which should be considered when interpreting its findings. First, only patients with stage T3N1M0 NPC were included in this study; future studies should consider patients with other disease stages. Second, CCRT plus adjuvant chemotherapy is also an optional treatment regimen in the advanced NPC [31], and the response to CCRT plus adjuvant chemotherapy remains to be investigated. Third, a nonparametric matching method was used to balance baseline characteristics between treatments in the retrospective study; however, selection bias between the two treatment regimens likely remained and may have affected the presented estimates; therefore, validation of the present findings in prospective trials is required.

In conclusion, we developed and validated a multi-task nomogram that could noninvasively predict the prognosis of T3N1M0 NPC patients with different treatment regimens and accordingly recommend an optimal treatment regimen in the multicentre cohorts. Our nomogram may act as a non-invasive and useful tool for promoting personalized treatment and optimizing management of NPC patients.

Declaration of Competing Interest

The authors have no conflicts of interest to disclose.