Background

Locally advanced head and neck squamous cell carcinomas (HNSCCs) are commonly treated with surgery and adjuvant radio(chemo)therapy or with definitive radiotherapy [1]. In definitive radiotherapy, outcome can be enhanced by concomitant chemotherapy [2]. However, overall survival (OS) and loco-regional control (LRC) still need to be improved. In this regard, recent data of the German Cancer Consortium Radiation Oncology Group (DKTK-ROG) report an LRC rate of 62.6% and OS of 59.6% after 2 years of follow-up in HNSCC [3].

To improve outcomes, much effort is made to establish personalised treatment strategies in radiation oncology [4, 5]. Precision medicine implies the potential to individualise therapy by the integration of multimodal data including genomics and radiomics.

A prominent publication investigated 440 computed tomography (CT)-based radiomic features including intensity, shape, texture and multiscale wavelet in lung cancer and HNSCC for prognostic value [6]. After training and validation, the best performing prognostic features of each category were identified. A worsened survival rate was associated with increasing radiomic heterogeneity. In addition, in one lung cancer cohort, gene expression profiles were correlated with radiomic features [6]. An association between upregulated cell cycle pathways and increased intra-tumour heterogeneity features (texture and wavelet feature) has been found.

In addition, with regard to immunotherapy as an upcoming therapeutic option in HNSCC, genetic prognosticators like the tumour mutational burden (TMB) are discussed [7] and some genetic variants might have predictive and prognostic value for therapy response. Tumour genome sequencing facilitates the determination of functional changes and risk groups of HNSCC patients [8]. The Cancer Genome Atlas (TCGA) characterised several frequent somatic variants in HNSCC including TP53 (cell cycle control and survival), KMT2D (chromatin modification) and FAT1 (Wnt/ß-catenin signalling, cell–cell contacts, cell orientation, cell fate) [9, 10], and the main signalling pathways in HNSCC are visualised [9]. As a cross-link to clinical features, variants in TP53 and FAT1 were predominantly found in human papillomavirus (HPV)-negative tumours [9]. In HPV-negative patients, variants in FAT1 were reported to be associated with beneficial outcome in surgically treated HNSCC patients [11]. The authors of this study found mutations in FAT1 as a strong, independent prognostic factor for overall survival in the TCGA cohort and could validate these findings in data of the International Cancer Genome Consortium (ICGC).

The aim of our study was to investigate radiomic tumour heterogeneity according to particular previously reported features and their associations with recurrent somatic driver mutations in HNSCC. With this hypothesis-driven approach of radiogenomic associations, we intended to find correlations that might refer to functional relationships or complemental characteristics of imaging features and genetic aberrations.

Methods

Patients and diagnostics

Twenty patients with locally advanced HNSCC were recruited for this prospective biomarker study. All declared their written informed consent. The study was approved by the local ethics committee (reference number 577/2014BO2) and conducted in accordance with the Helsinki Declaration. All patients were treated with definitive radiochemotherapy up to 70–77 Gy. HPV association was investigated by immunohistochemical staining for p16 or PCR-based assays. Clinical data was extracted from the medical reports.

Radiomics

Due to our limited cohort, we followed a hypothesis-driven approach for finding associations between radiomic heterogeneity and driver gene mutations. Based on the report by Aerts et al. [6], our first tested hypothesis postulated that cell cycle-related somatic mutations (i.e. driver gene mutations in TP53) might correspond with increased radiomic heterogeneity. As a second hypothesis, we investigated if other frequently mutated driver genes correlate with heterogeneity features of the tumour.

Based on our unenhanced planning CT scans (Somatom Sensation Open, Siemens Healthineers, Erlangen, Germany; slice thickness of 3 mm, in-plane pixel size of 1.27 mm, ordered subset expectation maximization [OSEM] 3D [4 iterations, 8 subsets] with a 3D Gaussian filtering for imaging reconstruction), we analysed the two best performing radiomic features for measuring intra-tumour heterogeneity that were described by Aerts et al. [6], namely “Run Length Nonuniformity” (Aerts et al.: Textural Feature 48) and “wavelet Grey Level Nonuniformity HLH” (Aerts et al.: Feature Group 4; decomposition of the image in mid-frequencies). Furthermore, we included “Grey Level Nonuniformity” (Aerts et al.: Textural Feature 47), as a complemental feature, as the authors also reported on this feature in the reference publication. Therefore, in total, three particular heterogeneity features were investigated, and the features were calculated following the previous report of Aerts et al. [6] for confirmability and standardisation.

The gross tumour volumes (GTVs) of the primary tumours were delineated for treatment planning by experienced radiation oncologists. These delineations were subsequently used for radiomic analyses. Due to concerns regarding the influence of dental artefacts [12], we investigated both the data of all 20 patients and, as a subgroup, the patients that had no CT artefacts in the area of interest. Texture features were preprocessed in a 3D fashion regardless of the in-plane, in-slice difference, and we categorised the intensity values in 64 different bins due to the sparse range of intensity values (between −250 to 120 Hounsfield units) across the GTV. For wavelet estimations we used the undecimated wavelet filter. If air or bony structures were included in the GTV, the delineations were adapted and extreme Hounsfield units were excluded for radiomics. Thereby, solely in one patient, the GTV was considerably modified due to massive air and bone involvement (oropharyngeal HNSCC with infiltration of the maxillary sinus).

Genetic analyses

Formalin-fixed paraffin-embedded (FFPE) tumour tissue (obtained at primary diagnosis) was provided by the pathology department and ethylenediaminetetraacetic acid (EDTA) blood samples were collected as normal tissue. We used a particular HNSCC cancer panel containing 327 genes which was originally designed by the DKTK-ROG partner site in Berlin. The library preparation and in-solution capture of the exonic regions were performed using the Agilent HaloplexHS technology (Agilent, Santa Clara, CA, USA). The samples were paired-end sequenced using the HiSeq2500 instrument (Illumina, San Diego, CA, USA). An in-house developed pipeline, called “megSAP”, was used for data analysis (version 0.1-755-g54185f9, https://github.com/imgag/megSAP). In brief, sequencing reads were aligned to the human genome reference sequence (GRCh37) using Burrows–Wheeler Aligner (BWA, version 0.7.17) [13]. Reads aligned to the same chromosomal position and with identical unique molecular identifiers were deduplicated by creating a consensus read constructed per position by choosing the most frequent base (with at least 75% frequency) or by replacing with “N”. Variants were called using Strelka2 (version 2.8.4) [14] and annotated with SnpEff/SnpSift (version 4.37) [15]. All variants were visually validated with the Integrative Genomics Viewer (version 2.3.97) [16], and quality control (QC) parameters were collected during all analysis steps [17]. For further interpretation, we uploaded all somatic variants to the Cancer Genome Interpreter (CGI) [18]. Somatic nucleotide variants were annotated as drivers based on the classification tier 1 and tier 2 (predicted driver mutations) for the single-nucleotide variants (SNVs).

Due to the limited cohort size, only mutations in the most recurrently mutated known driver genes with predicted driver variants according to the CGI database [18], namely TP53, FAT1 and KMT2D, were correlated with radiomic measures of tumour heterogeneity. One patient was excluded from genetic correlations since he showed a hypermutated genotype; thus, the functional impact of single variants remained unclear. For validity and clinical relevance, an allele frequency (AF) of ≥5% was required for reported mutations. However, we recorded driver mutations with lower frequency (<5%) in TP53, FAT1 and KMT2D, as the tumour content of some samples was comparably low. If these variants were annotated by the CGI database, the discrepancy between predicted function and low AF was considered debatable and therefore these variants were marked and excluded from analysis.

Statistics

For the statistical analyses, we used R [19] and SPSS (IBM Corp., Armonk, NY, USA). The Mann–Whitney U test and robust linear regression (M-estimator from MASS R package, log2-transformed values) were used for calculations. Significance estimations of regression coefficients were calculated by the robust F‑test (Wald test, sfsmisc R package). A p-value <0.05 was considered significant.

Results

Clinical data of the patients are summarized in Table 1. The most frequently detected driver gene variants and the HPV status are shown in Fig. 1. TP53, FAT1 and KMT2D were the most frequently mutated genes we found in our cohort. Therefore, we correlated the mutation status of these three genes with the image-based heterogeneity features.

Table 1 Patient characteristics
Full size table
Fig. 1
figure 1

Genetic profiles and according human papillomavirus (HPV) infection status of 19 patients are shown in the heatmap. Each column represents a different patient. One patient was excluded from genetic analysis due to a hypermutated genotype. The labelled (asterisk) TP53 and FAT1 mutations were below the 5% cut-off and therefore excluded from statistics. On the right side, the most frequently mutated driver genes are shown. On the y‑axis the frequency of the respective drivers is provided (%). At the top, the overall driver mutations are shown

Full size image

Regarding the three selected radiomic features indicating tumour heterogeneity, there was no significant correlation found with variants in TP53 or KMT2D. Therefore, the data are not shown. However, a significant association with FAT1 was found, as variants in FAT1 corresponded with reduced radiomic heterogeneity of the primary tumour (Grey Level Nonuniformity: p = 0.019; Fig. 2a; Run Length Nonuniformity: p = 0.046; Fig. 2b and wavelet Grey Level Nonuniformity HLH: p = 0.035; Fig. 2c). This association was found in all three selected heterogeneity features and the observation remained significant in two features when patients who had dental artefacts in the area of the primary tumour were excluded (Fig. 3). Two FAT1 mutations were excluded for reliability due to an AF <5% in one patient (marked with * in Fig. 1) and a hypermutated genotype in another patient (who was therefore excluded from genetic analysis). However, both of the variants were reproducible in the raw data. If these variants were included in the correlations, the association between FAT1 variants and image-based heterogeneity improved for all three features (Grey Level Nonuniformity: p = 0.005; Run Length Nonuniformity: p = 0.024 and wavelet Grey Level Nonuniformity HLH: p = 0.011, data not shown).

Fig. 2
figure 2

Association between FAT1 mutations and radiomic heterogeneity features. Correlations are shown for a Grey Level Nonuniformity, b Run Length Nonuniformity and c wavelet Grey Level Nonuniformity HLH

Full size image
Fig. 3
figure 3

Association between FAT1 mutations and radiomic heterogeneity features for patients without dental artefacts. Patients with dental artefacts in the area of interest (gross tumour volume) were excluded. Correlations are shown for a Grey Level Nonuniformity, b Run Length Nonuniformity and c wavelet Grey Level Nonuniformity HLH

Full size image

The association of somatic mutations in FAT1 and smaller GTVs of the primary tumour was not significant (p = 0.059; Fig. 4a). However, smaller GTVs of the primary tumour corresponded with reduced radiomic heterogeneity (Grey Level Nonuniformity: p < 0.001; Fig. 4b; Run Length Nonuniformity: p < 0.001; Fig. 4c and wavelet Grey Level Nonuniformity HLH: p < 0.001; Fig. 4d).

Fig. 4
figure 4

Gross tumour volume (GTV) and a associated FAT1 mutation status, b Grey Level Nonuniformity, c Run Length Nonuniformity and d wavelet Grey Level Nonuniformity HLH. For correlations between GTVs and radiomic heterogeneity features, a linear regression model was used and data are shown on a log2 scale

Full size image

Detailed information about the TP53, FAT1 and KMT2D variants in our cohort is visualized in the supplement (Suppl. Fig.).

Discussion

Imaging biomarkers and genetic variants are considered promising features to inform personalised therapeutic decisions. However, reports on correlations of radiomics and genomics remain sparse.

Regarding somatic mutations, our findings are well in line with previously reported mutation profiles in HNSCC [9]. For the investigation of correlations between radiomic data and somatic mutations, we used a hypothesis-driven approach. Our first hypothesis of a correlation between alterations in TP53 as a cell cycle regulator and increased tumour heterogeneity ascertained by radiomic features was not confirmed in our cohort. The previously described gene expression data indicating that increased activity of cell cycle pathways and enhanced proliferation correlate with tumour heterogeneity [6] did not translate into a significant association with somatic mutations of TP53. One could speculate that TP53 has broad effect on tumour development and treatment response, and a unidimensional correlation cannot be found.

For our second hypothesis, we investigated the association of tumour heterogeneity with two other frequently mutated driver genes of our cohort, FAT1 and KMT2D, which had also been identified as recurrently mutated in HNSCC in previous studies [9]. Interestingly, variants in FAT1 were associated with reduced tumour heterogeneity according to all three investigated radiomic heterogeneity features.

FAT1 was found to act as a tumour suppressor by binding ß‑catenin and subsequently decreasing ß‑catenin translocation to the nucleus [20]. Thereby, FAT1 indirectly inhibits cell proliferation and tumour growth. Inactivating mutations of FAT1 are therefore thought to promote the Wnt/ß-catenin signalling pathway [20]. Furthermore, FAT1 is linked to cell–cell contacts and seems to be required for tight cell–cell adhesions [20, 21] and cell polarity [21], as well as for control of cell migration [22] and invasion [22]. In addition, FAT1 contributes to the regulation of epithelial–mesenchymal transition (EMT) [23], which is thought to be associated with tumour aggressiveness. However, the role of FAT1 in tumourigenesis is discussed controversially as both tumour suppressive and oncogenic. These attributes might be dependent on different tumour types. The knockdown of FAT1 was found to reduce cell migration and invasiveness in oral squamous cell carcinoma, glioblastoma and colon cancer [22, 24, 25]. On the contrary, Hu et al. report on accelerated cell migration and EMT after FAT1 knockdown in oesophageal squamous carcinoma [26].

We found an association between FAT1-mutated tumours and reduced heterogeneity of the primary tumours according to radiomic features. Reduced heterogeneity corresponded to smaller primary tumour volumes, and in FAT1-mutated tumours, a trend towards reduced primary tumour volumes was observed, although significance levels were not reached. Postulating FAT1 to be a tumour suppressor in HNSCC [10], inactivating mutations would be expected to cause rather extensive volumes. However, the influence of FAT1 on proliferation in oral squamous cell carcinomas was described to be rather limited [22]. In this way, other oncogenes and tumour suppressors might have a comparably stronger influence on tumour growth and GTV extent.

As discussed above, inactivating/missense variants of FAT1 might result in reduced invasiveness, attenuated migration and looser cell–cell contacts. One could speculate that this translates into less radiomic heterogeneity and smaller tumour volumes as indicated in our cohort. As increased heterogeneity correlated with poor outcome [6] and smaller GTVs [27, 28] are associated with a good prognosis, our findings might support a recent publication reporting favourable outcomes of HPV-negative, surgically treated HNSCC patients, if they presented with mutant FAT1 [11]. Thus, our data suggest a possible interrelation between FAT1 mutations, reduced heterogeneity and smaller GTVs.

Our study revealed interesting preliminary findings in HNSCC patients. The limitation of our study is the small cohort and resulting limited effect size. Therefore, our radiogenomic observations remain solely descriptive and need to be confirmed in larger studies. A further limitation is the location of the biopsy. We sequenced FFPE material that was collected at primary diagnosis. The exact localisation within the primary tumour is therefore unknown. In case of intra-tumour genetic heterogeneity, some variants might be missed and others overestimated. However, we chose only the most frequently mutated driver genes for correlations. As these are well in line with previously reported drivers of HNSCC, these mutations are thought to determine relevant biological and functional variations that might translate into radiomic features.

Conclusion

The collection and integration of omic data for decision-making in precision medicine is essential. Furthermore, tumour characteristics can be investigated and correlated by radiogenomics. Here, we found that reduced radiomic tumour heterogeneity correlated with mutations in FAT1 and with smaller gross tumour volumes, possibly elucidating the cause of the previously described improved overall survival of HPV-negative, FAT1-mutated HNSCC patients.