Magnetic resonance imaging standardization for accurate grading of cerebral gliomas

Computer-aided diagnosis has attracted attention for the accurate grading of cerebral glioma. Most algorithms are only effective in relatively large datasets. Although multicenter data sharing is expanding, the results of cerebral glioma grading are not promising for multicenter data. Considering that multicenter images differ in contrast, we propose an effective image standardization method to reduce the disparity in image contrast of different datasets. The method is adopted in multiple sets of comparative experiments on a public dataset (BraTS2017) and a local dataset. The classification accuracy of experimental data relative to that of multicenter data without image normalization is improved by approximately 25% on average. Results demonstrate that the proposed approach is effective in solving the image contrast disparity of multicenter data. It also addresses the challenge of limited effective sample size in accurate cerebral glioma grading. The novel image standardization technology proposed in this work is a promising solution that can be integrated into expert systems.