Background: Glioma is the most common primary brain tumor in adults. Various machine learning tools via magnetic resonance imaging can make it a practical instrument in accurate and early diagnosis of tumors thereby assisting physicians in diverse diagnostic and therapeutic fields. The aim of this study is to automate the process of defining and determing the grade of glioma tumor with the use of a variety of learning algorithms.
Methods: This is a fundamental-applied study performed on multimodal MRI images of 285 patients with glioma tumors from the BraTS 2018 Challenge Database. In order to classify glioma tumors as high and low grade, first a was performed with U Net network for the definition purposes, then the results were incorporated for classification in VGG16 network to determine the exact grade of tumor.
Findings: The mean value of Dice Similarity Coefficient (DSC) for the classification designed for regions of the complete tumor, core of the tumor and the enhanced areas were 0.76, 0.70 and 0.71 respectively. The accuracy of the proposed classification based on VGG 16 network to determine the grade of tumor in both HGG and LGG groups was 99.01%.
Conclusion: Machine learning methods can he useful to determine the glioma tumor grade instead of using invasive proceedures like biopsy which in turn improves overall survival rate of these patients and their quality of life.