Computer-aided Identification Of Spinal Metastasis Originated From Lung Cancer And Prediction Of EGFR Gene Mutations

Objective: This study explored the establishment of a computer-assisted analysis method to automatically segmented spinal metastasis in magnetic resonance imaging(MRI)images of patients with spinal metastasis originated from lung cancer,and then further predict epidermal growth factor receptor(EGFR)gene mutation status,which has potential implications for treatment decisions and surgical options for patients with spinal metastasis.Methods: This study retrospectively included 130 patients with pathologically confirmed spinal metastasis from lung cancer(from January 2017 to September 2021).Stratified sampling was used to divide the training and validation sets by 8:2,with the presence or absence of mutations in the EGFR gene as the stratification criterion.This study was divided into two parts: the first part localized and segmented spinal metastases in MRI images,data preprocessing,model and training strategy design using the nnUnet framework.The results were evaluated with Dice coefficient,sensitivity,and specificity and so on.In the second part,the EGFR gene mutation status was predicted based on the automatically segmented spinal metastasis.Firstly,the radiomics features of the patient’s spinal metastasis were extracted,and features were selected using the Mann-Whitney U test,Least absolute shrinkage and selection operator(LASSO)and the Akaike’s information criterion(AIC)for radiomics feature dimensionality reduction.Finally,three machine learning classifiers(support vector machine,random forest and logistic regression)were constructed,and after selecting the classifier with the best performance,an Rad-Score model was constructed.The Nomogram model was established by fusing radiomics features and statistically significant clinical features.The performance of the model was evaluated using the area under the receiver operating characteristic curve(AUC),sensitivity,and specificity.Calibration curves and the Hosmer-Lemeshow test were used to evaluate model stability and goodness of fit.The clinical value of the model was evaluated by Decision Curve Analysis(DCA).Results: The first part of the results after automatic segmentation using the nnUnet framework showed that the Dice coefficient in the validation set reached 0.796,with a sensitivity of 0.852 and a specificity of 0.999.The second part of the results based on automatic segmentation of patients with spinal metastasis to predict EGFR gene mutation status showed that logistic regression achieved the best results among classifiers,with an AUC of 0.822,a sensitivity of 1.000 and a specificity of 0.539 in the validation set.The Rad-Score model based on logistic regression was developed.The Nomogram built by fusing smoking features achieved the best performance in predicting EGFR gene mutation status,with an AUC of 0.849,sensitivity of 0.769 and specificity of 0.769 in the validation set.Using the same extraction and dimensionality reduction methods,the radiomics features of manually segmented and automatically segmented spinal metastasis were compared,and four radiomics feature were found to have the same names.Conclusion: In this paper,we conducted a study of spinal metastasis originated from lung cancer based on nnUnet to solve the problem of automatic localization and segmentation of spinal metastasis in MRI.Then,we conducted a radiomics study to predict the mutation status of EGFR gene based on the automatically segmented spinal metastases,which provides a quantitative and non-invasive tool for clinicians.The combination of the two can solve current the problem that radiomics still needs to manually delineate lesions when assisting clinicians in decision-making,and has potential clinical application value.