Head and neck squamous-cell carcinoma (HNSCC) are tumours occurring in the oral cavity (OC), larynx, and pharynx. For males and females respectively, the estimated number of new cases in Europe for oral cavity cancers were 42,600 and 18,800; for larynx they were 36,000 and 3,900; and for pharynx they were 28,400 and 5,700. The treatment plan for HNSCC tumours is mostly performed through tumour, node, and metastasis (TNM) staging. This staging is dependent on the size of the primary tumour, its location, the extent of lymph nodes metastases, and the presence of distant metastasised nodes.

HPV-status has also been found to be prognostic for survival in oropharyngeal cancers, with HPV+ patients showing significantly higher locoregional control and lower mortality compared to HPV- patients. In the 8th edition of staging for oropharynx cancer, p16 status functions as an indicator of HPV status and has been included in staging.

Routine radiological imaging at presentation of cancer provides a source of information regarding the tumour volume that is currently not being used clinically. Advanced image analysis methods such as radiomics turn normal radiographic images into high-throughput data-mining formats, consisting of handcrafted features that describe tumour size, shape, and texture. Previous studies have shown that radiomics could play a role in improving prognosis of HNSCC. However, these studies require large amounts of data to be reliable, and effective models that are applicable to multiple centres in multiple countries, similar to staging, require data from various sources.

‘Big Data to Decide’ (BD2Decide) is a multi-centre European project that aims to improve clinical decision making in HNSCC through advanced analysis methods. The project has seven clinical centres participating in Italy (Parma, Brescia, and Milan), the Netherlands (Amsterdam and Maastricht) and Germany (Düsseldorf and Ulm). The multi-centric nature of the project provides the volume and diversity of data required for radiomics studies. Contrast-enhanced CTs from 837 patients were collected and split into retrospective (N=699) and prospective (N=138) cohorts, with gross tumour volume (GTV) delineated.

After pre-processing, radiomics features from four groups were extracted from the GTVs: first-order statistics, shape and size features, texture features, and wavelet/Laplacian of Gaussian (LoG) features. ComBat was used on the combined retrospective and prospective datasets, on a feature-by-feature basis. The batch effect removed from the dataset was the centre the image was collected from. Univariate and multivariate-cox proportional hazard modelling was done for feature selection and survival modelling respectively. Using the median cut-off for the predicted values of the retrospective cases, two risk groups are defined as high-risk and low-risk. A log-rank test was performed to determine the significance of this split, and the CI was used as a measure of modelling performance. Kaplan-Meier survival curves were used to depict the risk-groups.

We found that a four-feature radiomics signature (four filtered texture) can significantly group the validation cohort into two distinct risk groups (p < 0.05), with a CI of 0.68 (Figure 1). TNM-8 staging can determine two risk groups (p < 0.05), albeit with a higher CI of 0.70. However, the radiomics model does not take HPV status into account, and further analysis within subgroups may yield better results.

Figure 1: KM survival curves of the prospective cohort, stratified into low and high-risk groups by radiomics model signature scores.

Simon Andreas Keek is a PhD student at the D-Lab, Department of Precision Medicine, GROW- School for Oncology and Developmental Biology, at Maastricht University Medical Centre+ in Maastricht, the Netherlands.

Research Presentation Session
RPS 1405b Artificial intelligence and CT radiomics

An externally validated prognostic model based on CT radiomics to improve risk-stratification in head and neck cancer patients
S. A. Keek1, F. Wesseling1, L. Licitra2, K. Scheckenbach3, T. Hoffmann4, M. Ravanelli5, R. Leemans6, T. Poli7, P. Lambin1; 1Maastricht/NL, 2Milan/IT, 3Düsseldorf/DE, 4Ulm/DE, 5Brescia/IT, 6Amsterdam/NL, 7Parma/IT

Read the full abstract in the ECR 2020 Book of Abstracts
Keek SA, et al. (2020) An externally validated prognostic model based on CT radiomics to improve risk-stratification in head and neck cancer patients. Abstract RPS 1405b-1 in: ECR 2020 Book of Abstracts. Insights Imaging 11, 34 (2020). DOI 10.1186/s13244-020-00851-0