Comparison of two convolutional neural network models for automated classification of brain cancer types

Abstract

Convolutional neutral network (CNN) is widely used in the classification of types of brain cancer and many architectures of the CNN have been developed. Comparasions of various architectures on a specific clinical task is essential. This study aims to compare a deep transfer learning model with AlexNet and GoogleNet architectures for brain tumor classification on the T1-w magnetic resonance imaging (MR)I images. The comparison of the AlexNet and the GoogleNet architectures was implemented on the T1-w MRI images with three tumor types: glioma, meningioma and pituitary. The total images were 3,064 consisted of 1,426 gliomas, 708 meningiomas, and 930 pituitaries. 80% of datasets were for training and 20% of datasets were for testing. It is found that the accuracies for the AlexNet is 94.6% and for the GoogleNet is 92%. The sensitivity, specificity, precision and recall for the AlexNet are 94%, 95.2%, 94.6% and 46.9%, respectively. While sensitivity, specificity, precision and recall for the GoogleNet are 96.3%, 96.8%, 87.3% and 45.9%, respectively.