| Objective: This study explores the potential value of MRI radiomics in aiding the assessment of HER2 expression status in bone and brain metastases from breast cancer.Methods: This experiment was divided into three stages,110 patients with bone metastases and 67 patients with brain metastases originating from breast cancer were included from Liaoning Cancer Hospital in the first stage,the second stage predicts the HER2 expression status of bone metastases from breast cancer based on intratumoral heterogeneity segmentation,and the third stage predicts the HER2 expression status of brain metastases from breast cancer based on multi habitat analysis.For the MRI images of breast cancer bone metastases,the Kmeans unsupervised clustering algorithm based on local entropy was used to segment the interior of the bone metastases into different regions;for the MRI images of breast cancer brain metastases,the active,necrotic and oedema region of the brain metastases were marked into different regions by manual outlining.After that,radiomics features were calculated and extracted for each region separately,and the most valuable features were screened by Mann-Whitney U test and LASSO algorithm to build radiomics model,model performance was evaluated with AUC,specificity,and sensitivity as evaluation indexes,and the prediction performance of three machine learning classifiers: Logistic regression,Support vector machine and Random forest were compared.Finally,the Nomogram model was established by multi-region fusion and multi-habitat fusion,and the application value of the Nomogram model was evaluated by calibration curve and decision curve.Results: For bone metastases from breast cancer,we divided the tumors into two subregions based on intratumoral heterogeneity,and constructed the whole tumor area model(WTA)and two subregion models(Subregion＿1,Subregion＿2)by the best classifier with Logistic regression,where the AUC of the training set were 0.796,0.810,and 0.824 in order,and the AUC of the test set were 0.701,0.767,0.734 in order,the results showed that the prediction performance of Subregion＿1 and Subregion＿2 was better than that of WTA,and the Nomogram model fusing Subregion＿1 and Subregion＿2 had higher AUC values compared with the single subregion model.For brain metastases from breast cancer,the active and oedema region models for T1 CE sequences as well as T2 W sequences(T1CE＿TAA,T1CE＿POA,T2W＿TAA,T2W＿POA)were constructed based on the best classifier with Logistic regression,where the AUC of the training set were 0.810,0.805,0.839,0.834 in order,and the AUC of the test set were 0.762,0.751,0.801,0.734 in order,the results showed that the prediction performance of the active region model was better than that of the oedema region model,while the Nomogram model fusing T2W＿TAA and T2W＿POA had higher AUC values compared with the single habitat model.Conclusion: The results of this study suggest the importance of combining intratumoral heterogeneity to predict HER2 expression status in breast cancer bone metastases,multiple habitats of breast cancer brain metastases show variability in the prediction of HER2 expression status,and that the prediction effect of tumoral active area is better than that of peritumoral oedema area. |
