Quantitative magnetic resonance imaging radiomics predicts photoreceptorness status in retinoblastoma.

BACKGROUND

Molecular characteristics of retinoblastoma cannot be assessed before treatment because tumor biopsy is contraindicated. Photoreceptorness reflects photoreceptor-related gene expression and tumor differentiation. Non-invasive imaging biomarkers that capture this biology are therefore needed.

OBJECTIVE

To evaluate whether quantitative radiomics derived from pretreatment magnetic resonance imaging can predict loss of photoreceptorness in retinoblastoma and validate this approach in an independent cohort.

MATERIALS AND

METHODS

In this retrospective multicenter study, patients with retinoblastoma who underwent primary enucleation and had both pretreatment T2-weighted magnetic resonance imaging and genome-wide messenger RNA expression data were included. Tumors in the highest and lowest photoreceptorness quartiles were analyzed. Whole-tumor segmentations were used to extract radiomic features with PyRadiomics. Multiple machine-learning pipelines were evaluated using repeated stratified cross-validation, and the best-performing model was tested in an independent cohort.

RESULTS

Forty-five patients (median age, 18 months [range, 2-70], 18 female) were included: 29 in the training cohort and 16 in the independent testing cohort. The best-performing model used recursive feature elimination with a random forest classifier and achieved a mean cross-validated area under the receiver operating characteristic curve of 0.83 in the training cohort. In the independent testing cohort, the model achieved an area under the receiver operating characteristic curve of 0.81 (95% confidence interval, 0.54-1.00) for predicting loss of photoreceptorness.

CONCLUSION

Quantitative magnetic resonance imaging radiomics showed preliminary moderate-to-good discriminatory performance for non-invasively predicting photoreceptorness status in retinoblastoma. These proof-of-concept findings suggest that radiomics may capture imaging features related to molecular tumor differentiation and could support the development of personalized treatment strategies.