Multi-sequence MRI-based Radiomics For Predicting Metastaes Origins Of Brain Metastasis

Objective: Accurate and non-invasive identification of the origin of brain metastases is important for formulating personalized treatment for patients with brain metastases as soon as possible.Based on multi-sequence brain magnetic resonance imaging(MRI),this study was performed to explore the value of radiomics in predicting the origins of brain metastases.Methods: This study was divided into two experiments.The first experiment enrolled156 patients with brain metastases,including 70 from non-small cell lung cancer(NSCLC),63 from breast cancer,and 23 from other cancers.All patients underwent T1 CE and T2 W MRI scans.The radiomics features were extracted from different regions of brain metastases and selected using the Mann-Whitney U test,least absolute shrinkage and selection operator(LASSO)regression and Akaike’s information criterion(AIC).The logistic regression classifier was applied to construct the radiomics models based on the single and multi-region of brain metastases.The second experiment enrolled 140 patients,including 60 from NSCLC,60 from breast cancer,and 20 from other origins.Using the patient-level and population-level clustering,brain metastases were divided based on heterogeneity.The Res Net50 with a global average pooling layer(RN-GAP)was proposed to calculate deep learning-based features.The open source package“Pyradiomics” based on the Python platform was applied to calculate handcrafted radiomics features.The method of feature selection and model construction was same as the first experiment.The single regional models were constructed based on each subregion and the whole tumor region.The combined radiomics models were constructed based on the combination of handcrafted radiomics features and deep learning features.The receiver operating characteristic curve(ROC)was used to evaluate the prediction performance of each model.Results: The results of the first experiment showed that the combined models based on the multi-region and the single-region model based on Brain-tumor interface(BTI)showed high prediction performance.The developed multi-region combined radiomics models generated good AUCs for identifying the NSCLC origin and breast cancer origin of brain metastases,0.827 and 0.871 in the training cohort,0.801 and 0.813 in the test cohort.The single-region model based on BTI generated good AUCs for identifying the NSCLC origin and breast cancer origin of brain metastases,0.913 and 0.870 in the training cohort,0.913 and 0.870 in the test cohort.The results of the second experiment showed that the marginal subregion S1 hold higher predictive value than the inner subregion S2.The combined radiomics model with deep learning features had better prediction effect.The AUCs of the combined radiomics model for predicting NSCLC origin and breast cancer origin was 0.860 and 0.909 in the training cohort,and 0.819 and0.872 in the test cohort.Conclusion: Brain MRI-based radiomics can accurately predict the origin of brain metastases,which is a potential tool to assist radiologist to determine the origin of brain metastases and guide personalized treatment for patients with brain metastasis.The radiomics models based on different regions of brain metastases are valuable for predicting the origin of brain metastases.The intratumoral segmentation and deep learning methods are of great significance for identifying the source of brain metastases.