An artificial neural network model for glioma grading using image information

OBJECTIVE: To explore the feasibility and efficacy of artificial neural network for differentiating high-grade glioma and low-grade glioma using image information.
 Methods: A total of 130 glioma patients with confirmed pathological diagnosis were selected retrospectively from 2012 to 2017. Forty one imaging features were extracted from each subjects based on 2-dimension magnetic resonance T1 weighted imaging with contrast-enhancement. An artificial neural network model was created and optimized according to the performance of feature selection. The training dataset was randomly selected half of the whole dataset, and the other half dataset was used to verify the performance of the neural network for glioma grading. The training-verification process was repeated for 100 times and the performance was averaged.
 Results: A total of 5 imaging features were selected as the ultimate input features for the neural network. The mean accuracy of the neural network for glioma grading was 90.32%, with a mean sensitivity at 87.86% and a mean specificity at 92.49%. The area under the curve of receiver operating characteristic curve was 0.9486.
 Conclusion: As a technique of artificial intelligence, neural network can reach a relatively high accuracy for the grading of glioma and provide a non-invasive and promising computer-aided diagnostic process for the pre-operative grading of glioma.