Risk-free WHO grading of astrocytoma using convolutional neural networks from MRI images

Abstract

Astrocytoma is the most common and aggressive brain tumor, in its highest grade, the prognosis is ‘low survival rate’. Spinal tap and biopsy are the methods executed in order to determine the grade of astrocytoma. Once the grade of astrocytoma is determined, treatment is planned to improve the life expectancy of oncological subjects. Spinal tap and biopsy are invasive diagnostic procedures. Magnetic resonance imaging (MRI) being widely used imaging modality to detect brain tumors, produces the large volume of MRI data each moment in clinical environments. Automated and reliable methods of astrocytoma grading from the analysis of MRI images are required as an alternative to biopsy and spinal tape. However, obtaining molecular information of brain cells using non-invasive methods is challenging. In this research work, an automatic method of astrocytoma grading using Convolutional Neural Networks (CNN) has been proposed. Results have been validated on a locally developed dataset, obtained from Department of Radiology (Diagnostics), Bahawal Victoria Hospital, Bahawalpur, Pakistan. The proposed method proved a significant achievement in terms of accuracy as 99.06% (for astrocytoma of Grade-I), 94.01% (for astrocytoma of Grade-II), 95.31% (for astrocytoma of Grade-III), 97.85% (for astrocytoma of Grade-IV), and overall accuracy of 96.56%.