Radiomics in the differential diagnosis of focal brain lesions: a retrospective study

INTRODUCTION: Glioblastoma and solitary metastases are the most common malignant neoplasms of the brain, characterized by high mortality and severe disability in patients. The method of choice for neuroimaging glioblastomas and metastases is contrast-enhanced magnetic resonance imaging. However, differentiation between the two is often difficult due to similar radiological features on MRI. Radiomics and machine learning can differentiate the primary origin of brain metastases and identify pathological tumor types noninvasively. OBJECTIVE: Application of texture analysis for differential diagnosis of glioblastomas and metastases of different etiologies. MATERIALS AND METHODS: 169 MRI studies from the RSCRR database were used in the study, 11 of which visualized morphologically differentiated glioblastoma of the brain, 55 lung cancer metastases and 103 breast cancer metastases. Segmentation of the regions of interest was performed semi-automatically in the free 3D-Slicer software with the ability to upload radiomic features from the regions of interest. For each lesion, 107 radiomic features were calculated from T1 and T2 sequences. Statistics: The calculation of statistical indicators was performed in a computer program for statistical data processing IBMSPSS Statistics 23. In statistical data processing, the Mann-Whitney statistical criterion for quantitative indicators and correlation analysis using the Pearson criterion were used to reduce the feature space. The reduction of the feature space and the selection of predictors by the feature_importances measure based on decision forests were carried out. Machine learning models were built in Python 3.10 using specialized libraries. RESULTS: For the model based on radiomic features extracted from T1 sequence, random forest showed the most efficient result, ROC-AUC=0.815 [0.749; 0.874]. For the model based on the radiomic features extracted from the T2 sequence, random forest showed the most effective result, ROC-AUC=0.817 [0.743; 0.873]. For the complex model based on radiomic features extracted from T1 and T2 sequences, random forest showed the most efficient result, ROC-AUC=0.855 [0.789; 0.906]. DISCUSSION: The classification models and their metrics obtained by us indicate that the radiomic features extracted from T2 weighted MR images make it possible to differentiate breast cancer metastases from lung cancer metastases with higher sensitivity than the features extracted from T1 weighted MR images. We also identified a large number of significantly different indicators in the construction of models for the differentiation of glioblastomas from metastases, which demonstrates the prospects of this direction. It is planned to continue the study with the expansion of samples. Our conclusions are also confirmed by the research results of our foreign colleagues. CONCLUSION: The models we have obtained are highly accurate and sensitive to the differentiation of metastases of various etiologies and demonstrate significant potential in continuing this study with an expansion of samples.