Correlation of gene expression with magnetic resonance imaging features of retinoblastoma: a multi-center radiogenomics validation study

MRI and gene expression data were collected from (1) Emory Eye Center, Atlanta, Georgia, USA; (2) Amsterdam UMC, location VUmc, the Netherlands; and (3) Essen University Hospital, Essen, Germany. Patients were included if there was (1) histopathologic diagnosis of retinoblastoma; (2) pre-treatment MRI available including at least T1-, T2-, and T1 contrast-enhanced sequences; and (3) genome-wide gene expression data. The extreme outliers for the expression studies that drastically diverged from their corresponding cohort were excluded, as these may signify low mRNA profiling quality and/or a high percentage of non-tumor material. The hierarchical clustering analysis with WARD2 method and based on the normalized expression values was performed. Subsequently, samples that widely branched out main clusters and sub-branches of the cohorts were removed from the analysis. In total, 2 samples from Amsterdam and 4 samples from Atlanta were designated as outliers by their clustering patterns and subsequently discarded. Seventeen patients were previously reported in a study on genetic markers for high-risk retinoblastoma [16], while the current study reports on associations between MRI features and gene-expression profiles.

A previously validated imaging atlas with definitions and examples of features was adopted for scoring [6]. This atlas was composed by radiologists with experience in retinoblastoma MRI from the European Retinoblastoma Imaging Collaboration and compiled 25 MRI features that may be encountered in retinoblastoma imaging. The imaging features that were assessed included features informing on tumor morphology (e.g., shape, definition of tumor margin), intra-tumoral findings (e.g., calcifications, proportion of enhancement), and intra-ocular findings (e.g., number of tumor lesions, vitreous seeding). Two radiologists with expertise in retinoblastoma imaging (M.C.d.J. and P.d.G., with 10 and 17 years of experience, respectively) individually scored the features blinded for patient data. Inter-reader agreement was calculated using Cohen’s kappa. Subsequently, discrepancies were resolved by consensus and the consensus scores were adopted for association with gene expression values. The MRI units included 1.5-T (n = 19, 53%) and 3-T (n = 17, 47%) systems, with an average pixel size of 0.34 mm (range 0.27–0.90 mm), an average section thickness of 0.90 mm (range 0.6–3 mm), and an average intersection gap of 0.75 mm (range 0.3–3.6 mm).

Within each gene expression platform, the photoreceptorness score per sample was calculated as the averaged normalized expression of the predefined list of 2753 photoreceptorness genes [7]. The photoreceptorness scores were scaled by calculating the Z-score. Approximate sample photoreceptorness relative to the mean Z-scores from study sites were combined and associated with patient-matched MRI feature scores using Spearman’s rank correlation coefficient or the Kruskal–Wallis test in accordance with the test study, unless when lack of data points did not permit association analysis (i.e., less than 3 alternative scores per feature). Correlation analysis was performed within R, considering p < 0.05 significant.

The contrast design between the test and validation studies was not comparable, impeding repeated analysis with a relatively comparable design matrix and statistical power. To address the issue, two analyses were performed. Firstly, sample relationships were examined by the differentially expressed genes within MRI feature in the previous study (test): number of lesions (n = 70 genes), tumor location (n = 355 genes), subretinal seeding (n = 873 genes), and vitreous seeding (n = 19 genes). The normalized expressions of the genes previously identified as differentially expressed for features were extracted in the current sample (validation set). This was followed by performing k-mean clustering (k = 3) and principal component analysis (PCA). The partitioning outcomes and the first two principal components of expression were then overlaid and plotted to examine the relationship between samples with different MRI feature scores.

Secondly, the differential expression analysis within each site was performed when at least three samples at each side of the contrast design were available, permitting a minimum reliable statistical power. Subsequently, the end results of the analysis were merged, and gene ontology enrichment analysis was performed by Toppgene (https://​toppgene.​cchmc.​org/​).

Thirty-six retinoblastoma patients from three retinoblastoma referral centers in Atlanta, USA (n = 17), Amsterdam, Netherlands (n = 13), and Essen, Germany (n = 6) were included in this study. Of the patients, 14 were female (42%), 34 (94%) were unilateral, mean age at scan date was 24 months (SD 18 months), and mean MRI examination year was 2012 (SD 6 years, range 2002–2018). The findings within this set (hereafter referred to as “validation set”) were compared with the analysis of the “test set” for which patient demographics were previously described [6]. Cohen’s kappa for inter-reader agreement for adopted imaging features was on average 0.42 (moderate agreement), ranging from 0 to 0.68 (Appendix Table 1).

Similar to the findings in the test set, the photoreceptorness gene expression score showed a gradual distribution among samples from the three centers; examples of a low and a high photoreceptorness case were provided (Fig. 1). Three MRI features that significantly correlated to the photoreceptorness gene expression score in the test set were also found to be correlated with photoreceptorness in the current validation set: number of lesions (multiple lesions showing lower photoreceptorness, p = 0.01), growth pattern (diffuse growth showing lower photoreceptorness, p = 0.03), and tumor location (tumor in the entire globe or entire retina showing lower photoreceptorness, p = 0.02) (Table 1). For these MRI features, the differential expression of photoreceptorness was remarkably similar for the feature categories (Fig. 2) in the validation set compared with the test set. The number of lesions showed an inverse linear relationship with photoreceptorness in the validation set, with multifocality showing low photoreceptorness similar to findings from the test set. Similarly, diffuse tumor growth was again found to be correlated with low photoreceptorness. In both the test set and the validation set, tumors within the entire retina or filling the entire globe showed low photoreceptorness expression, while tumors with greater components behind the equator exhibited higher photoreceptorness expression. Therefore, the validated MRI phenotype of low photoreceptorness retinoblastoma consisted of multifocal, diffuse-growing retinoblastoma filling the globe or growing along the entire retina (Fig. 2). For this imaging phenotype, combined radiology and histopathology figures were presented for low photoreceptorness cases (Fig. 3). Histopathology assessment of the three cases with the lowest photoreceptorness from the Amsterdam center showed lack of rosettes (particularly Flexner-Wintersteiner rosettes), indicating poor differentiation (Fig. 3) [25, 26]. Other imaging-photoreceptorness correlations from the test set were not validated: eye size and increased choroidal enhancement beneath the tumor were significantly associated with photoreceptorness in the test set, but not in the validation set. Statistically significant findings associated with lower photoreceptorness only in the current validation set included plaque-shaped tumors (p = 0.01) and tumors with irregular margins (p = 0.01) (Appendix Fig. 1).

This validation study partially validated the test study for differentially expressed genes based on MRI features. K-Means clustering (k = 3) based on the genes previously identified but with their expression levels in the current cohort again revealed clustering of imaging traits (Fig. 4). When clustering was based on differentially expressed genes for tumor location in the test set (n = 355), tumors located in entire globe or retina congregated in the validation set (cluster 1). Similarly, tumors central/posterior in the retina congregated in cluster 3. For the differentially expressed genes based on the number of lesions in the test set (n = 70), there was a close relationship between samples harboring 1 and 1–5 lesions in the validation set (cluster 1), while cluster 3 almost exclusively included tumors with 5–10 lesions or > 10 lesions. Regarding the other two imaging features associated with many differentially expressed genes in the previous study—subretinal and vitreous seeding—no distinctive clustering pattern was found in the current study (Appendix Fig. 2). Gene ontology enrichment for differentially expressed genes per imaging feature identified in the current validation study did not show overlap with finding from the test study (Appendix Table 2). Additionally, previous imaging features associated with SERTAD3 and KAL1 genes were not validated in this study.