The prognostic value of CT radiomic features from primary tumours and pathological lymph
nodes in head and neck cancer patients
Zhai, Tiantian
DOI:10.33612/diss.111448998
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date: 2020
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Zhai, T. (2020). The prognostic value of CT radiomic features from primary tumours and pathological lymph nodes in head and neck cancer patients. University of Groningen. https://doi.org/10.33612/diss.111448998
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
CHAPTER 5
External validation of nodal failure prediction
models including radiomics in head and neck
cancer
Manuscript in preparation Tian-Tian Zhaia,b,*, Frederik Wesselingc, Johannes A. Langendijkb, Zhenwei Shic, Petros
Kalendralisc, Lisanne van Dijkb, Frank Hoebersc, Roel J.H.M. Steenbakkersb, Andre
Dekkerc, Leonard Weec, Nanna M. Sijtsemab a Department of Radiation Oncology, Cancer Hospital of Shantou University Medical
College, Shantou, China
b Department of Radiation Oncology, University of Groningen, University Medical
Center Groningen, Groningen, The Netherlands
c Department of Radiation Oncology (MAASTRO), GROW School for Oncology and
Development Biology, Maastricht University Medical Centre+, Maastricht, The Netherlands *Corresponding author Tian-Tian Zhai and Frederik Wesseling contributed equally to this work.
Abstract Purpose
To externally validate the previously developed pre-treatment prediction models for lymph nodes failure after definitive radiotherapy in head and neck squamous cell carcinoma (HNSCC) patients.
Materials and Methods
This external validation cohort consisted of 143 HNSCC patients treated between July 2007 and June 2016 by curative radiotherapy with or without either cisplatin or cetuximab. Imaging and pathology reports during follow-up were analyzed to indicate persisting or recurring nodes. The previously established clinical, radiomic and combined models were validated on this cohort by assessing the concordance index (c-index) and model calibration.
Results
113 patients with 374 pLNs were suitable for final analysis. There were 20 (5.3%) nodal failures from 15 patients after a median follow-up of 36.1 months. Baseline characteristics and radiomic features were comparable to the training cohort.
Both the radiomic model (Least-axis-length of lymph node (LALLN) and correlation of gray level co-occurrence matrix (Corre-GLCM)) and the combined model (T stage, gender, WHO performance score, LALLN and Corre-GLCM) showed good agreement between predicted and observed nodal control probabilities. The radiomic (c-index: 0.71; 95% confidence interval (CI): 0.84) and combined (c-index: 0.71; 95% CI: 0.59-0.82) models performed better than the clinical model (c-index: 0.57; 95% CI: 0.47-0.68) on this cohort, with a significant difference between the combined and clinical models (z-score test: p=0.005).
Conclusion
The combined model including clinical and radiomic features was externally validated and proved useful to predict nodal failures and could be helpful to guide treatment choices before and after curative radiation treatment for node positive HNSCC patients.
5
Introduction
Lymph node management after curative radiotherapy with or without systemic treatment for nodal positive head and neck squamous cell carcinoma (N+ HNSCC) is challenging. Not all clinical or radiographic persisting nodes will contain viable tumour cells and responding nodes still could recur during follow-up [1–4]. Therefore, decisions for salvage treatment based on physical examination or radiographic response on imaging could lead to over- or under-treatment [3–7].
In current practice, nodal response after radiotherapy is commonly based on neck ultrasound (US) with or without fine needle aspiration (FNA), computed tomography (CT), positron emission tomography (PET) or magnetic resonance imaging (MRI). However, predictive performances are variable and depend on several factors such as cut-off values and radiologist's experience [8–12]. In general, negative predictive values (NPVs) of radiological modalities are quite high, leading to safe conservative management in cases with radiological complete response. In contrast, positive predictive values (PPVs) are generally low, resulting in unnecessary neck dissections in case of incomplete radiologic response [1,2,6,7].
Three models were developed in Chapter 4 to identify pathological lymph nodes (pLNs) that are at risk to persist or recur after definitive radiotherapy with or without systemic treatment in HNSCC patients. The three models used clinical variables, radiomics, and a combination of clinical variables and radiomics as input parameters. The models were trained and internally validated on 277 HNSCC patients with 1022 pathological lymph nodes (pLNs) from a single center (UMC Groningen, UMCG).
The model based on radiomics showed a better predictive performance than the model with clinical factors. Performance improved further when both models were combined. The radiomic model consisted of two parameters: the least axis length of the lymph node (LALLN) representing the size of a node and the correlation of grey levels within a lymph node (Corre-GLCM) describing the heterogeneity of a lymph node. Those parameters extracted from pre-treatment CT-images of individual lymph nodes performed better than nodal prediction based on N-stage only.
The main goal of this study was to validate the models developed in Chapter 4 in a large and independent external cohort. External validation of the models could achieve
a transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) type 4 level of evidence [13], which is the first step in making these models available for supporting treatment decisions for HNSCC patients. Ideally, individual nodal prediction could lead to radiation dose intensification or direct surgical dissection for high-risk nodes, or in case of recurrence to super-selective neck dissection, to limit morbidity of salvage surgery [5,12,14].
Materials and Methods Study population
The study population included 143 N+ HNSCC patients treated with curative intent at the department of radiation oncology in Maastricht, the Netherlands (MAASTRO) from July 2007 to June 2016. Institutional research ethics board approval was obtained and the need for written consent was waived for this retrospective study. All patients had HNSCC originating in the oral cavity, pharynx or larynx. They were primarily treated with definitive radiotherapy (RT) (68-70 Gy, 2 Gy per fraction) with or without either cisplatin or cetuximab. Patients were eligible only if they had contrast-enhanced CT-based RT planning data available and standard follow-up procedure.Patients were excluded if neck surgery was performed before radiotherapy or metallic CT artefacts were seen in the region of pathological lymph nodes. Figure 1 details 113 patients eligible in the study and the reasons for exclusion.
Clinical parameters
Patient characteristics were collected in the same way as described in the model development. T-stage (T3-T4 vs. T1-T2), N-stage (N2-N3 vs. N1), and clinical stage (IV vs. III) were scored according to the 7th edition of the American Joint Committee on Cancer Staging Manual [15]. Gender (male vs. female), WHO performance-status (WHO PS; 1-3 vs. 0), treatment modality (radiotherapy only vs. radiotherapy with systemic treatment) and age at diagnosis were derived from the medical records. Human papillomavirus (HPV) status of oropharyngeal cancer (HPV- vs. HPV+) was assessed by immunohistochemistry of the cellular protein p16.
CT image acquisition, pathological lymph node segmentation and radiomic features
5
neck according to standard clinical contrast enhanced scanning protocols with Siemens CT scanner (Somatom Sensation Open, Siemens, Forchheim, Germany). The CT pixel spacing was 0.98 × 0.98 × 3.0 mm3.
To ensure consistent methodology for external validation, the definition of pathological lymph nodes (pLNs) and radiomic features reported by Zhai et al. were used [16,17]. All pLNs were incorporated in the high-risk treatment volume and received a radiation dose of 68-70 Gy. All pLNs were individually contoured on the planning CT scans by T.Z. and F.W. using a treatment planning system (Eclipse™, Varian, Palo Alto, CA, USA). Overall 82 radiomic features per lymph node were extracted using Matlab software (version R2014a; Mathworks, Natick, USA) according to the image biomarker standardisation initiative (IBSI) [18].
Endpoints
The endpoint was the cumulative incidence of nodal failure. The events of nodal failure were defined as cytologically or histopathologically proven persistent or recurrent nodal disease within or overlapping with the originally treated lymph node area at a minimum of 12 weeks after treatment. All patients’ follow-up records and imaging were thoroughly checked by T. Z. and F.W. according to the definition of endpoints reported in Chapter 4.
Data analysis
Patient characteristics of the MAASTRO cohort were compared to the UMCG training and validation cohorts by using a chi-squared test for categorical variables and independent sample t-test for normal distributed variables. Boxplots of raw values of radiomic features in the models (least-axis-length of lymph node (LALLN) and correlation of gray level co-occurrence matrix (Corre-GLCM)) were generated on UMCG and MAASTRO cohorts to compare the distribution of radiomic features.
The prognostic scores were calculated using the following clinical model, radiomic model and combined model equations :
Prognostic score_Clinical model (1):
2.08 * Gender + 1.08 * T-stage + 0.85 * N-stage + 0.78 * WHO PS Prognostic score_Radiomic model (2):
0.78 * LALLN + 3.14 * Corre-GLCM Prognostic score_Combined model (3):
1.88 * Gender + 0.93 * T-stage + 0.91 * WHO PS + 0.71 * LALLN + 2.44 * Corre-GLCM In the models, Gender = 0 if female, 1 if male; T-stage = 0 if T1-T2, 1 if T3-T4; N-stage = 0 if N1, 1 if N2-N3; WHO PS = 0 if 0, 1 if 1-3; LALLN = raw value in centimeters; Corre-GLCM = raw value.
Patients were divided into low-, intermediate- and high-risk groups by using the lower and upper quantile of the prognostic scores from the UMCG training cohort (Supplementary A).
Model discrimination and calibration
The models’ discriminative power was assessed by Harrell’s concordance index (c-index). Kaplan-Meier curves were used to show the differences between different risk groups and were compared using log-rank tests. The Z-score test was used to compare the c-indexes.
The model calibration was checked by following the standard test procedure proposed by Royston et al [19]. The baseline cumulative hazard function H0(t) was described in the simplified look-up table in supplementary B. The predicted survival function of S(t,PS) = exp(-H0(t))exp(PS) was used to calculate the curve for each lymph node [19]. The expected curve in each group was determined by averaging the curves of all lymph nodes of that group. The expected and observed curves were superimposed to assess how well they fit.
The model discrimination is considered preserved in the validation cohort, if the calibration slope β is not significantly different from 1.
Two tailed p-values < 0.05 were considered statistically significant. Statistical analysis was performed using the R software (version 3.2.1).
Results
In total, 113 patients with 374 pLNs were eligible for this study (Figure 1). The largest number of patient exclusions (n=15) was due to neck surgery before radiotherapy (Figure 1, Supplementary C). A total of 20 nodal failures (5.3%) from 15 patients (13.3%) after a median follow-up period of 36.1 months (range: 1.2-108.5) were observed. Two lymph node failures that occurred after 5 years were excluded from the analysis, because recurrences after this period are less common and more likely to be part of
5
second primary malignancy [20,21]. The baseline characteristics of the MAASTRO cohort in comparison to the UMCG training and validation cohorts are summarized in Table 1. There was no significant difference between the datasets from the two institutes
Figure 1. Recruitment pathway for patients in this study.
Abbreviations: HNSCC= head and neck squamous cell carcinoma; ROI = region of interest (pathological lymph nodes).
Figure 2 Boxplots of the radiomic features in the UMCG training, UMCG validation and Maastro cohorts. Abbreviations: LALLN = least-axis-length of lymph node; GLCM = gray level of co-occurrence matrix.
except for treatment modality and WHO PS. The MAASTRO cohort comprised fewer patients treated with chemotherapy or cetuximab and more patients with WHO PS of 1 compared to UMCG cohort.
The boxplots of the two radiomic features are shown in Figure 2. The median values and distributions of the LALLN and Corre-GLCM were similar for the three cohorts.
Table 1 Baseline characteristics of head and neck cancer patients in the UMCG training, validation and MAASTRO cohorts.
Characteristic
UMCG
training cohort validation cohortUMCG MAASTRO cohort p-Value* n=165 % n=112 % n=113 % Age at diagnosis (median ± SD, years) 60 ± 9 64 ± 10 62 ± 8 0.983b Gender 0.894c Female 45 27.3 37 33.0 30 26.5 Male 120 72.7 75 67.0 83 73.5 T-stagea 0.409c T1 11 6.7 13 11.6 13 11.5 T2 35 21.2 22 19.6 28 24.8 T3 49 29.7 22 19.6 30 26.5 T4 70 42.4 55 49.1 42 37.2 N-stagea 0.745c N1 33 20.0 19 17.0 22 19.5 N2 123 74.5 91 81.3 87 77.0 N3 9 5.5 2 1.8 4 3.5 Clinical stagea 0.709c III 25 15.2 13 11.6 19 16.8 IV 140 84.8 99 88.4 94 83.2 Treatment modality <0.001c RT only 49 29.7 39 34.8 66 58.4
RT with systemic treatment 116 70.3 73 65.2 47 41.6
WHO PS <0.001c
0 105 63.6 65 58.0 35 31.0
1 52 31.5 35 31.3 73 64.6
2 7 4.2 10 8.9 4 3.5
5
The clinical model showed a c-index of 0.57 (95% confidence interval (CI): 0.47-0.68) in the MAASTRO cohort. The radiomic and combined models performed better than clinical model with c-indexes of 0.71 (95% CI: 0.59-0.84; z-score test: p=0.052) and 0.71 (95% CI: 0.59-0.82; z-score test: p=0.005), respectively (Table 2).
The Kaplan-Meier curves are shown for the 3 models in Figure 3. They show that the clinical model was not able to stratify the lymph nodes into different risk groups (p-value = 0.498). The nodal control probability prediction using both the radiomic and combined models showed significant differences between the three risk groups with p-values of 0.003 and < 0.001, respectively. Furthermore, the differences between two risk groups
Table 1 Baseline characteristics of head and neck cancer patients in the UMCG training, validation and MAASTRO cohorts-continued.
Characteristic
UMCG
training cohort validation cohortUMCG MAASTRO cohort p-Value* n=165 % n=112 % n=113 % Tumor site 0.573c oral cavity 11 6.7 11 9.8 6 5.3 oropharynx 75 45.5 70 62.5 57 50.4 nasopharynx 4 2.4 5 4.5 0 0.0 hypopharynx 33 20.0 11 9.8 21 18.6 larynx 42 25.5 15 13.4 29 25.7 HPV status 0.201c OPC negative 48 29.1 30 26.8 27 23.9 OPC positive 25 15.2 34 30.4 28 24.8 OPC unknown 2 1.2 6 5.4 2 1.8 Not OPC 90 54.5 42 37.5 56 49.6 No. of patients
with nodal failures 37 22.4 19 17.0 15 13.3
No. of pLNs 558 467 374
No. of nodal failures 68 12.2 23 4.9 20 5.3 Abbreviations: T = tumor; N = lymph node; RT = radiotherapy; WHO PS = World Health Organization performance status; HPV = human papillomavirus; OPC= oropharyngeal cancer; pLNs = pathological lymph nodes.
* p-Value: UMCG training cohort vs. MAASTRO cohort a According to the 7th edition of the AJCC/UICC staging system b p-Value was calculated using the independent t-test c p-Value was calculated using the chi-square test
Table 2 Discrimination (c-index) and calibration slope (β) evaluated by clinical, radiomic and combined model in MAASTRO cohort.
Clinical model Radiomic model Combined model
Harrell c-index 0.570 0.712 0.707
Harrell c-index (95 % CI) 0.463-0.675 0.586-0.838 0.594-0.820
Calibration slope (β) 0.309 1.044 0.762
Calibration slope (p-Value) 0.011 0.865 0.294
(low-risk vs. intermediate-risk and intermediate-risk vs. high-risk) are compared and shown in Supplementary D. The radiomic model showed non-significant discrimination between the low-risk group and the intermediate-risk group (p = 0.103), and separated the high risk group from the intermediate risk group with a p-value of 0.062. The combined model showed a significant separation between all risk groups (low-risk vs. intermediate-risk: p = 0.048, intermediate-risk vs. high-risk: p = 0.016, and low-risk vs. high-risk: p = 0.001 respectively), indicating better discrimination than the radiomic model.
The calibration slopes of the radiomic and combined models were not significantly different from 1 (Table 2), indicating preservation of the discriminative value of the models in the external validation cohort.
The averaged predicted curves for different risk groups were superimposed on the observed curves in Figure 3. The clinical model consistently under-predicts the nodal control probabilities in all risk groups, and the largest deviation between predicted and observed nodal control was observed in the high-risk group. The predicted curves by the radiomic model demonstrated good agreement with observed curves for all risk groups. The combined model showed good calibration for all risk groups within the first 2 years’ follow-up. However, the nodal control probability after 2 years was a little underestimated for the high-risk group.
Discussion
Three pre-treatment prediction models developed in Chapter 4 showed the relationship between clinical variables, radiomic features and nodal failure rate in HNSCC patients [. These models make the risk stratification of pathological lymph nodes possible and accordingly provide the opportunity of personalized decision making targeted on the
5
Figur e 3 Ob ser ved and pr edict ed nodal c on trol pr obabilities f or lo w, in termediate and high risk gr
oup
s in MAAS
TR
O c
ohort.
individual lymph nodes. To our knowledge, this is the first TRIPOD type 4 study to validate and compare nodal failure prediction models in an external patient cohort. Using exactly the same risk factors and endpoints, the clinical model did not enable significant discrimination between the risk groups, while radiomic and combined models showed good performance in terms of calibration and discrimination in the independent MAASTRO cohort.
patients with bulky lymph nodes were excluded from the analysis because of their policy to perform neck dissection for advanced nodal stage before start of radiotherapy. This could be the explanation for the lower nodal failure rate in the MAASTRO cohort than that in UMCG cohort (5.3% vs. 8.9%), which might have affected the ability of the clinical model to detect a significant result. The MAASTRO cohort comprised more patients treated without chemotherapy or cetuximab, possibly because of worse WHO performance status and more patients older than 70 years that were not considered candidate for systemic treatment, in line with the MACH-NC meta-analysis [22]. Those variations between cohorts may contribute to the lack of external validity of the clinical model [23]. Moreover, although the clinical model performed better in the UMCG validation cohort than that in the MAASTRO cohort, there was also c-index reduction in the UMCG validation cohort, which had similar patient characteristics to the UMCG training cohort . This illustrates that the clinical model based on patient-specific features may be less robust for nodal failure prediction.
Clinical variables such as total nodal volume and N-stage have been shown as prognostic factors for regional control in multiple studies [24–28]. But, N-stage in the clinical model refers to the number and location of pathological lymph nodes within the patient, but does not provide information on each separate lymph node. Clinical variables take certain aspects of all lymph nodes into consideration, and thus they seem more suitable for failure prediction of the total neck region.
For individual lymph node failure prediction, the features of each individual lymph node are needed. Parameters such as extra-nodal extension (ENE), reported as a significant risk factor for nodal control can be assessed on the level of every lymph node [29]. However, for the non-surgical HNSCC patients treated primarily with radiotherapy, pathologic ENE is not available and clinical ENE assessment based on CT-images alone has been shown to be unreliable [30].
The radiomic features extracted from each lymph node are expected to be more informative, objective and quantitative measures of the phenotypes of lymph nodes than clinical variables. Firstly, all radiomic features are objectively calculated using formulas while some clinical variables such as WHO PS are subjective parameters. Secondly, radiomic features are quantitative and continuous variables, which might
5
be more informative than categorical clinical features such as N-stage and presence or absence of central necrosis. Thirdly, radiomics provides non-invasive features that take the whole lymph node into account, which could be an advantage compared to the assessment of pathological ENE which is based on limited needle biopsy samples. For these rationales, we expect that radiomic features could outperform clinical variables in predicting individual lymph node failure [31–33].
In the radiomic model, LALLN representing the nodal size and Corre-GLCM representing the heterogeneity of lymph nodes were identified as the most important independent prognostic variables for nodal control. Despite a series of studies investigating the relationship between nodal size and regional control [2,17,24,25], few studies have observed nodal sizes’ impact on the control of each separate node [27] . This study confirmed that the size of a specific pathological lymph node is a negative prognostic variable for its nodal control. The size feature of LALLN showed strong correlation with and performed better than lymph node volume reported by Vergeer, et al. [27]. It is worth mentioning that a round-shaped lymph node tends to have a larger LALLN than an oval-shaped lymph node with similar volume, indicating that LALLN includes information not only on the size, but also on the shape of lymph nodes.
The value of the texture feature Corre-GLCM was high when the lymph node had more heterogeneous CT-intensities. In our opinion, the heterogeneous intensities may indicate intra-nodal necrosis status [17], as the lymph nodes with necrosis showed increased values of Corre-GLCM. In this cohort, the lymph node with higher Corre-GLCM values had a worse nodal control, which is in line with the results reported in other studies [27,34,35].
The radiomic model showed comparable discrimination with the combined model (Table 2), and both were well calibrated (Figure 3). The question arises as to whether the combined model is improved by adding clinical features to the radiomic model. First, the difference between the low- and intermediate-risk groups was not significant (log-rank test: p = 0.103) by using the radiomic model, but became significant (p = 0.048) by including clinical variables (supplementary D). This indicates that clinical variables contribute to the low risk lymph node prediction, which can also be seen in Kaplan-Meier curves of the clinical model (Figure 3). Moreover, the separations between
low-, intermediate- and high-risk group curves are larger if lymph nodes are stratified by the combined model than by the radiomic model. Furthermore, the predicted and observed curves of the combined model showed good calibration within 2 years after the treatment, and a small underestimation of the nodal control probability after 2 years for high-risk group. In this study, the predictive performance in the first two years is considered more important since nodal failures of patients treated primarily with radiotherapy mainly occur within 2 years [20,36]. The underestimation of the nodal control probability for the high risk group could be explained by the large number of censoring data after 2 years follow-up [19]. These results indicate that the combined model might be more applicable to clinical practice. For cohorts which have large deviations in clinical variables compared to our study cohort, the radiomic model could be suggested since the radiomic model also had good performance.
The combined prognostic model achieved satisfactory discrimination and calibration in this external validation study. This model showed the ability to classify pathological lymph nodes into high- or low-risk groups prior to the treatment. Therefore, the neck management could be optimized for these selected lymph nodes. As a potential future strategy, we would recommend that lymph nodes stratified into low-risk group according to the model, could be followed with a wait and see policy instead of surgical dissection in case of clinical persisting disease shortly after radiotherapy. Intensified radiation or surgical dissection could be suggested for those lymph nodes with high failure risks aiming at improving nodal control probability. This hypothesis should be investigated within the setting of prospective studies, and useful post-treatment surveillance such as PET-CT should be considered as a complement to this pre-treatment prediction model [37,38]. To further improve this model, we think that advanced imaging modalities and artificial intelligence will play an increasingly important role [33]. In addition, the accurate assessment of biomarkers such as ENE, which has been proven to be a prognostic factor for regional control in a series of studies and was included in the 8th AJCC staging, could potentially improve the nodal failure prediction[29,30,39].
Conclusion
In conclusion, this TRIPOD type 4 external validation study evaluated and compared three prognostic models for nodal failure in HNSCC patients. The combined model that
5
included both clinical and radiomics features, was well calibrated and enabled significant discrimination between different risk groups in the independent cohort, and could be used to guide decisions on customized treatment strategies targeted on individual lymph nodes.
References
1. Studer G, Huber GF, Holz E, Glanzmann C. Less may be more: nodal treatment in neck positive head neck cancer patients. Eur Arch Oto-Rhino-Laryngology 2016;273:1549–56.
2. van den Bovenkamp K, Dorgelo B, Noordhuis MG, van der Laan BFAM, van der Vegt B, Bijl HP, et al. Viable tumor in salvage neck dissections in head and neck cancer: Relation with initial treatment, change of lymph node size and human papillomavirus. Oral Oncol 2018;77:131–6.
3. Thariat J, Ang KK, Allen PK, Ahamad A, Williams MD, Myers JN, et al. Prediction of neck dissection requirement after definitive radiotherapy for head-and-neck squamous cell carcinoma. Int J Radiat Oncol Biol Phys 2012;82:e367–74.
4. Goenka A, Morris LGT, Rao SS, Wolden SL, Wong RJ, Kraus DH, et al. Long-term regional control in the observed neck following definitive chemoradiation for node-positive oropharyngeal squamous cell cancer. Int J Cancer 2013;133:1214–21. 5. Brizel DM, Prosnitz RG, Hunter S, Fisher SR, Clough RL, Downey MA, et al. Necessity
for adjuvant neck dissection in setting of concurrent chemoradiation for advanced head-and-neck cancer. Int J Radiat Oncol Biol Phys 2004;58:1418–23.
6. Donatelli-Lassig AA, Duffy SA, Fowler KE, Ronis DL, Chepeha DB, Terrell JE. The effect of neck dissection on quality of life after chemoradiation. Otolaryngol - Head Neck Surg 2008;139:511–8.
7. Machtay M, Moughan J, Trotti A, Garden AS, Weber RS, Cooper JS, et al. Factors associated with severe late toxicity after concurrent chemoradiation for locally advanced head and neck cancer: An RTOG analysis. J Clin Oncol 2008;26:3582–9. 8. Hermann RM, Christiansen H, Rödel RM. Lymph node positive head and neck
carcinoma after curative radiochemotherapy: A long lasting debate on elective post-therapeutic neck dissections comes to a conclusion. Cancer/Radiotherapie 2013;17:323–31.
9. Pellini R, Manciocco V, Turri-Zanoni M, Vidiri A, Sanguineti G, Marucci L, et al. Planned neck dissection after chemoradiotherapy in advanced oropharyngeal squamous cell cancer: The role of US, MRI and FDG-PET/TC scans to assess residual neck disease. J Cranio-Maxillofacial Surg 2014;42:1834–9.
5
10. King AD, Yu KH, Mo FKF, Law BKH, Yuen TWC, Bhatia KS, et al. Cervical nodal metastases from head and neck squamous cell carcinoma: MRI criteria for treatment assessment. Head Neck 2016;38:E1598–604.
11. Folkert MR, Oh JH, Setton J, Apte AP, Thorstad WL, Schoder H, et al. Predictive Modeling of Outcomes Following Definitive Chemoradiation Therapy for Oropharyngeal Cancer Based on FDG-PET Image Characteristics. Int J Radiat Oncol 2013;87:s3.
12. Liauw SL, Mancuso AA, Amdur RJ, Morris CG, Villaret DB, Werning JW, et al. Postradiotherapy neck dissection for lymph node-positive head and neck cancer: The use of computed tomography to manage the neck. J Clin Oncol 2006;24:1421–7. 13. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD Statement. Eur Radiol 2015;67:1142–51.
14. Robbins KT, Shannon K, Vieira F. Superselective Neck Dissection After Chemoradiation. Arch Otolaryngol Head Neck Surg 2007;133:486–9.
15. Pfister DG, Spencer SA, Brizel DM, Burtness BA, Busse PM, Caudell JJ, et al. Head and neck cancers, version 1.2015 featured updates to the NCCN guidelines. JNCCN J Natl Compr Cancer Netw 2015;13:847–56.
16. Zhai TT, van Dijk L V., Huang BT, Lin ZX, Ribeiro CO, Brouwer CL, et al. Improving the prediction of overall survival for head and neck cancer patients using image biomarkers in combination with clinical parameters. Radiother Oncol 2017;124:256– 62.
17. Zhai T-T, Langendijk JA, van Dijk L V., Halmos GB, Witjes MJH, Oosting SF, et al. The prognostic value of CT-based image-biomarkers for head and neck cancer patients treated with definitive (chemo-)radiation. Oral Oncol 2019;95:178–86.
18. Alex Zwanenburg, Stefan Leger, Martin Vallières SL. Image biomarker standardisation initiative 2016. doi:10.17195/candat.2016.08.1.
19. Royston P, Altman DG. External validation of a Cox prognostic model: Principles and methods. BMC Med Res Methodol 2013;13:33.
20. Pagh A, Grau C, Overgaard J. Failure pattern and salvage treatment after radical treatment of head and neck cancer. Acta Oncol (Madr) 2016;55:625–32.
21. Morris LGT, Sikora AG, Patel SG, Hayes RB, Ganly I. Second primary cancers after an index head and neck cancer: Subsite-specific trends in the era of human papillomavirus - Associated oropharyngeal cancer. J Clin Oncol 2011;29:739–46. 22. Pignon JP, Maître A le, Maillard E, Bourhis J. Meta-analysis of chemotherapy in
head and neck cancer (MACH-NC): An update on 93 randomised trials and 17,346 patients. Radiother Oncol 2009;92:4–14.
23. Mallett S, Royston P, Waters R, Dutton S, Altman DG. Reporting performance of prognostic models in cancer: A review. BMC Med 2010;8.
24. Rietbergen MM, Witte BI, Velazquez ER, Snijders PJF, Bloemena E, Speel EJ, et al. Different prognostic models for different patient populations: Validation of a new prognostic model for patients with oropharyngeal cancer in Western Europe. Br J Cancer 2015;112:1733–6.
25. Dua B, Chufal KS, Bhatnagar A, Thakwani A. Nodal volume as a prognostic factor in locally advanced head and neck cancer : Identifying candidates for elective neck dissection after chemoradiation with IGRT from a single institutional prospective series from the Indian subcontinent. Oral Oncol 2018;87:179–85.
26. Regueiro CA, Aragón G, Millán L, Valcárcel FJ, de la Torre A, Magallón R. Prognostic factors for local control, regional control and survival in oropharyngeal squamous cell carcinoma. Eur J Cancer 1994;30:2060–7.
27. Vergeer MR, Doornaert P, René Leemans C, Buter J, Slotman BJ, Langendijk JA. Control of nodal metastases in squamous cell head and neck cancer treated by radiation therapy or chemoradiation. Radiother Oncol 2006;79:39–44.
28. Vallières M, Kay-Rivest E, Perrin LJ, Liem X, Furstoss C, Aerts HJWL, et al. Radiomics strategies for risk assessment of tumour failure in head-and-neck cancer. Sci Rep 2017;7.
29. Mermod M, Tolstonog G, Simon C, Monnier Y. Extracapsular spread in head and neck squamous cell carcinoma: A systematic review and meta-analysis. Oral Oncol 2016;62:60–71.
30. Chai RL, Rath TJ, Johnson JT, Ferris RL, Kubicek GJ, Duvvuri U, et al. Accuracy of computed tomography in the prediction of extracapsular spread of lymph node metastases in squamous cell carcinoma of the head and neck. JAMA Otolaryngol -
5
Head Neck Surg 2013;139:1187–94.
31. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, De Jong EEC, Van Timmeren J, et al. Radiomics: The bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749–62.
32. Aerts HJWL. The potential of radiomic-based phenotyping in precisionmedicine a review. JAMA Oncol 2016;2:1636–42.
33. Liu Z, Wang S, Dong D, Wei J, Fang C, Zhou X, et al. The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges. Theranostics 2019;9:1303–22.
34. Lin Y-C, Chen S-W, Hsieh T-C, Yen K-Y, Yang S-N, Wang Y-C, et al. Risk Stratification of Metastatic Neck Nodes by CT and PET in Patients with Head and Neck Cancer Receiving Definitive Radiotherapy. J Nucl Med 2015;56:183–9.
35. Haralick RM, Dinstein I, Shanmugam K. Textural Features for Image Classification. IEEE Trans Syst Man Cybern 1973;3:610–21.
36. Johansen S, Norman MH, Dale E, Amdal CD, Furre T, Malinen E, et al. Patterns of local-regional recurrence after conformal and intensity-modulated radiotherapy for head and neck cancer. Radiat Oncol 2017;12:87.
37. Hisham Mehanna, Wai-Lup Wong, Christopher C. McConkey, Joy K. Rahman, Max Robinson, Andrew G.J. Hartley, Christopher Nutting, Ned Powell, Hoda Al-Booz, Martin Robinson, Elizabeth Junor, Mohammed Rizwanullah the P-NTMG. PET-CT surveillance versus neck dissection in advanced head and neck cancer. N Engl J Med 2016;374:1444–54.
38. Mehanna HM, Wong WL, McConkey CC, Rahman JK, Robinson M, Hartley AGJ, et al. PET-NECK: A multi-centre, randomized, phase III, controlled trial (RCT) comparing PETCT guided active surveillance with planned neck dissection (ND) for locally advanced (N2/N3) nodal metastases (LANM) in patients with head and neck squamous cell cancer. J Clin Oncol 2019;33:6009–6009.
39. Huang SH, O’Sullivan B. Overview of the 8th Edition TNM Classification for Head and Neck Cancer. Curr Treat Options Oncol 2017;18.
Supplementary A
Table 1 Lower and upper quantiles of the clinical, radiomic and combined models’ prognostic scores in UMCG training cohort
Clinical model Radiomic model Combined model
Lower quantiles 1.864 1.583 3.322
Upper quantiles 3.718 2.403 5.266
Supplementary B
Table 1 Baseline cumulative hazard of clinical, radiomic and combined models
Months after treatment
Baseline cumulative hazard
Clinical model Radiomic model Combined model
0 0.001908681 0.003218882 0.000215524 3 0.001908681 0.003218882 0.000215524 6 0.002673073 0.004570504 0.000315566 9 0.003085375 0.005228812 0.00036995 12 0.003534623 0.005875717 0.000431622 15 0.00362134 0.00599388 0.00044363 18 0.003714926 0.006117587 0.000456139 21 0.004108399 0.006637696 0.000510902 24 0.004232831 0.006782059 0.000528258 27 0.004378426 0.006938877 0.000548225 30 0.004540466 0.007103782 0.000569835 33 0.004720419 0.007291072 0.000593874 36 0.004720419 0.007291072 0.000593874 39 0.004720419 0.007291072 0.000593874 42 0.004720419 0.007291072 0.000593874 45 0.004720419 0.007291072 0.000593874 48 0.004720419 0.007291072 0.000593874 51 0.004720419 0.007291072 0.000593874 54 0.004720419 0.007291072 0.000593874 57 0.004720419 0.007291072 0.000593874 60 0.004720419 0.007291072 0.000593874
5
Supplementary C
Table 1 Clinical stage of the 15 patients excluded as the neck dissection before radiotherapy.
Characteristic n=15 % T-stage T1 4 26.7 T2 9 60.0 T3 1 6.7 T4 1 6.7 N-stage N1 2 13.3 N2 8 53.3 N3 5 33.3 Clinical stage III 2 13.3 IV 13 86.7 Supplementary D
Table 1 Pairwise comparisons between risk groups using Log-Rank test (p values)
Clinical model Radiomic model Combined model
p-value Low Intermediate Low Intermediate Low Intermediate
Intermediate 0.500 0.103 0.048
