The future of contemporary neuroimaging does not solely lie in novel image-capturing technologies, but also in better methods for extraction of useful information from these images. Scientists see great promise in radiomics, i.e. the methodology for analysis of multiple features in medical image. However, there are certain issues in this field impairing reproducibility of results. One such issue is no standards in establishing the regions of interest.

OBJECTIVE

To introduce a standardized method for identification of regions of interest when analyzing MR images using radiomics; to test the hypothesis that this approach is effective for distinguishing different histological types of gliomas.

MATERIAL AND METHODS

We analyzed preoperative MR data in 83 adults with various gliomas (WHO classification, 2016), i.e. oligodendroglioma, anaplastic oligodendroglioma, anaplastic astrocytoma, and glioblastoma. Radiomic features were computed for T1, T1-enhanced, T2 and T2-FLAIR modalities in four standardized volumetric regions of interest by 356 voxels (46.93 mm³): 1) contrast enhancement; 2) edema-infiltration; 3) area adjacent to edema-infiltration; 4) reference area in contralateral hemisphere. Subsequently, mathematical models were trained to classify MR-images of glioma depending on histological type and quantitative features.

RESULTS

Mean accuracy of differential diagnosis of 4 histological types of gliomas in experiments with machine learning was 81.6%, mean accuracy of identification of tumor types — from 94.1% to 99.5%. The best results were obtained using support vector machines and random forest model.

CONCLUSION

In a pilot study, the proposed standardization of regions of interest demonstrated high effectiveness for MR-based differential diagnosis of oligodendroglioma, anaplastic oligodendroglioma, anaplastic astrocytoma and glioblastoma. There are grounds for applying and improving this methodology in further studies.