Recently the texture analysis has been applied to help diagnose and predict the clinical outcome of various tumor types. To investigate the diagnostic ability of radiomics-based machine learning in differentiating atypical low-grade astrocytoma (LGA) from anaplastic astrocytoma (AA).
The present study included 106 patients diagnosed with LGA and AA at the department of neurosurgery, West China Hospital from January-2014 to December-2018. The radiomics features of tumor tissues were extracted from contrast-enhanced T1 weighted imaging (T1C) taken prior to treatment. Nine diagnostic models were established with three selection methods and three classification algorithms including Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and random forest (RF). The sensitivity, specificity, accuracy, and area under curve (AUC) of each model were calculated, based on which the diagnostic ability of each model was evaluated.
The algorithms-based models presented the most remarkable diagnostic performances. As for LDA-based models, the optimal one was the combination of LASSO+LDA with an AUC of 0.802. Furthermore, LASSO+SVM represented the optimal SVM-based model with an AUC of 0.799. As for RF-based models, LASSO+RF and GBDT+RF demonstrated feasible discriminative ability with AUC of 0.778.
The radiomic-based machine learning could potentially serve as differentiating tool for atypical low-grade astrocytoma and anaplastic astrocytoma. In view of the present study and previous reports, contrast-enhanced computed tomography-based radiomics can be recommended in the clinical practice.
The authors.
Has not received any funding.
All authors have declared no conflicts of interest.